Questor · Unbounded Optimization

What if exact, interactive tradeoff optimization—of complex systems in the face of conflicting objectives—were available in any domain, on modest compute, faster than everything else, and produced results consumable and actionable by decision makers, both human and AI?

The answer

That day has arrived.

Questor is the enabler—system optimization and refinement, always at the ready. Precise, rapid, and available exactly when your systems need it most. See how it works in the video below.

The Plan a Trip panel from the Questor travel demo: choose your allowed flights, hotels and cars, pick your objectives, then hit Optimize.
The Travel optimization · 2 min 59 sec, with sound
Speed · Completeness

Faster. More actionable.

Faster where best-of-breed approaches cope. Complete where they can’t.

Questor and best-of-breed exact optimization approaches were given identical exact-Pareto optimization problems across multiple objectives. All had to return every optimal system on the frontier — not a sample of it. As the problem grows, the difference stops being how long it takes and becomes whether it finishes at all.

PARETO FRONTIER 5,114 0.238 s Questor · complete frontier Dynamic Programming 56.2 s · complete frontier Integer Linear Programming 150 s Few to no results. 0 60s 120s 180s
PARETO FRONTIER 11,197 0.551 s Questor · complete frontier Dynamic Programming 180 s No result. Integer Linear Programming Failed on a smaller problem 0 60s 120s 180s
PARETO FRONTIER 79,532 12.2 s Questor · complete frontier Dynamic Programming Failed on a smaller problem Integer Linear Programming Failed on a smaller problem 0 60s 120s 180s
Explore the Engine Read the Thesis
Applications

Questor is domain-independent.

Wherever systems are built from many components and subsystems—then viewed, compared, and selected against competing objectives—Questor drives the production of systems that achieve the optimal tradeoffs across every material aspect. Especially where objectives conflict and a system that is ideal in all respects is rare, or absent entirely.

A montage of the domains Questor optimizes: travel, IT and networks, financial portfolios, industrial processes, architecture, recruitment, healthcare, gene expression, nutrition and more.
Every decision · complex systems · better outcomes

Apply it horizontally

Screenshots of four example technology demonstrations — step through them.

1 of 4 · Travel 2 of 4 · Healthcare 3 of 4 · Nutrition 4 of 4 · Generative AI

Travel itineraries

Every option considered, so you don’t have to.

Example optimization problem

Explores all possible itineraries across flights, accommodations and local transportation options.

Maximize Comfort Minimize Connections Minimize Days Off Preferred Date Minimize Local Travel Time Minimize Time-of-Day Deviation Minimize Total Cost Minimize Total Flight Time
Red-eye flight On-site gym On-site parking 24-hour check-in Rental car
The Travel demo after optimizing: five components and six objectives over 9,953,280 possible itineraries, resolved in 0.655 seconds into five tradeoff paths for 684 optimal itineraries. Path A reaches Premium comfort with no stops at $3,634; Path E reaches $639 by accepting two days off preferred.
See the full interfaceHide the full interface
The complete Travel screen: destination and departure choices, five component selectors, ten objectives, the results table of five tradeoff paths, the Valid Choice Navigator narrowing Path D Set 382 to one exact itinerary, and the resulting bookable trip itinerary totalling $751 over 14h 50m of flight time.

Healthcare & medications

Optimize regimens across material clinical risks.

Example optimization problem

Explores all regimens of medications indicated for the patient’s diagnoses.

Minimize orthostatic hypotension Minimize constipation Minimize back pain Minimize dizziness Minimize insomnia Minimize fatigue Minimize nausea Minimize diarrhea Minimize anemia
16-Med Regimen — the optimization summary and the resolved tradeoff table.
See the full interfaceHide the full interface
The complete 16-Med Regimen screen: 16 medications and five objectives over 718,525,815,480,000,000 possible combinations, resolved in 0.003 seconds into five tradeoff paths and 12 optimal regimens.

Nutrition

Built from what’s actually on the menu.

Example optimization problem

Buildable bowls at a build-your-own restaurant.

Minimize Calories Maximize Protein Minimize Sodium
Vegan Vegetarian Paleo
The Build a Bowl demo after optimizing: eight ingredient slots and three nutrition objectives over 173,606,602,752 possible bowls, resolved in 12.445 seconds into five tradeoff paths and 8,297 optimal bowls.

Optimize Generative AI

Let a model find what exists. Questor constructs optimal systems using what exists. And improves the LLM’s performance as well.

Example optimization problem

Questor optimizes the products generated by LLMs. But it also optimizes the performance of an LLM by optimizing Model Serving Configurations. LLM parameters optimized include model × precision × context, accelerator × tensor parallelism, batching, speculative decoding, serving engine, replicas, KV-cache policy, autoscaling, and placement. Exact optimization is essential as nearly half of Pareto optimal Model Serving Configurations are unreachable via approximate methods.

Maximize Resilience Minimize Ops complexity Minimize Cost ($/hr) Maximize VRAM headroom Minimize p95 Latency Minimize Power draw Maximize Quality (eval) Maximize Throughput
The Model Serving Configuration demo after optimizing: ten decisions and eight objectives over 460,800,000 possible deployments, resolved in 6.830 seconds, with the Tradeoff Map showing seven paths from a starting deployment.

Other applications

The same engine, the same exact method—across domains we haven’t built a technology demo for yet. Happy to work with you to build a Proof of Concept: provide your own categories and objectives.

Categories are the parts to choose from—medications, bowl ingredients, flights. Objectives are the measures to minimize or maximize across them.

Some examples include: Matchmaking, Recruitment Candidates, IT Configurations & Networks, Financial Portfolios, Industrial Processes, Architectural Specifications, and Gene Expression.

Explore themHide them

Matchmaking

Optimize couples’ tradeoffs across personal and relational attributes.

Example optimization problem

Prospective couples drawn from a matchmaking site that maximize overall compatibility, are closest geographically and maximize each members’ key criteria.

Recruitment Candidates

Teams optimized for the jobs and tasks.

Example optimization problem

Teams of recruits on LinkedIn, Monster, Indeed and Dice that have the needed skills in my profile, the most years of experience for these skills, are the closest to our headquarters and are likely to get along based on interests and social media.

IT Configurations & Networks

Balance latency, security and cost across the stack.

Example optimization problem

Compositions of network micro-protocols that maximize bandwidth and hardware compatibility but minimize latency, jitter and energy consumption.

Combinations of motherboards, CPUs, controllers and GPUs maximize memory bandwidth and clock rate while minimizing cost and delivery time with NVIDIA as the preferred GPU.

Financial Portfolios

Allocate for return without overreaching on risk.

Example optimization problem

Portfolios of financial instruments that make the optimal tradeoff across return, risk, volatility and my preferred countries of record.

Industrial Processes

Push throughput without burning energy or quality.

Example optimization problem

Combinations, for cold rolled steel, of surface preparations, primers, color coats and topcoats that maximize weather resistance, minimize application time and have a total cost less than the industry standard process.

Architectural Specifications

Systems of components that optimally meet the requirements and specifications.

Example optimization problem

Best options for whole-house water treatment, either centralized, point-of-use, or hybrid, considering reverse osmosis, filtration and distillation, or combining these modalities, for a five-bedroom house, having the lowest cost, highest capacity and minimum maintenance.

Gene Expression

Analysis beyond hierarchical clustering.

Example optimization problem

Sets of target genes, from gene expression microarray data, that express to the probes together, and maximally so.

The evidence

Improved tradeoffs. Better outcomes. Lower cost.

The founders of Questor, working at its Surveyor Health spinoff, proved with high statistical significance that managing and improving medication risk tradeoffs reduces the Total Cost of Care by nearly $1 out of every $5 spent on high-risk patients.1 This was achieved by human pharmacists employing Questor’s evolutionary ancestor, SurveyorAI, which computed and visualized risk tradeoffs making them actionable by clinicians.

−19.3%

Total cost of care

IEHP, 2018–2019, with disease-management pharmacists on the platform.

Peer-reviewed · JMCP 2021

$6.4M

Saved per 1,000 high-risk patients, annually

The same reduction expressed per capita, so a plan can scale it to its own population.

Derived from the JMCP study

−62%

Emergency-department utilization

At the largest FQHC, now in its fourth year of daily use.

Deployment data · Surveyor Health

45–54%

Fewer heart-failure admits and readmits

Across a large Southern California hospital system.

Deployment data · Surveyor Health

Four years of daily use. Not a demonstration and not a retrospective — the platform has run in clinics every working day since 2018.

Journal of Managed Care & Specialty Pharmacy nameplate. Published study title page: 'Economic and utilization outcomes of medication management at a large Medicaid plan with disease management pharmacists using a novel artificial intelligence platform from 2018 to 2019', authored in part by Erick Von Schweber and Linda Von Schweber.

1 Published in JMCP · September 2021 · Vol. 27, No. 9

Read the peer-reviewed study in JMCP

To scale from Surveyor Health out to serve the millions of high-risk patients in the U.S., we’re now integrating a spoken conversational agent, named Aimi — sponsored by Amazon AWS and co-developed with Caylent, a Premier AWS Partner and a member of Anthropic’s Claude Partner Network — along with Questor to enable Closed-Loop Collaborative Care Intelligence. This will enable the outcomes observed in the JMCP clinical study to be deployed far and wide at a scale the pharmacist workforce cannot reach. Large systems are asking for this precisely because they cannot hire their way to millions of high-risk patients.

Who it’s for

Who Questor adds value to.

Generative AI & Big Tech

A new class of AI processing beyond training and inference—a competitive differentiator that builds on existing LLMs. Improves LLM performance by optimizing Model Serving Configurations.

Relevant to

Frontier AI labs · LLM & foundation-model providers · AI answer engines · hyperscale cloud platforms

Organizations That Live With the Tradeoff

The buyer who carries the cost of a good-enough answer every day—where a configuration has to satisfy objectives that genuinely conflict, and nobody can see all the options at once.

Relevant to

Health plans & hospitals · airlines & travel · industrial process owners · financial portfolios · IT & network architects · engineering firms

Healthcare Organizations & Health IT Vendors

Easy medication optimization brought right into the EHR, at the point of care for primary-care providers.

Relevant to

Health plans · hospitals · clinics · emergency departments · physician groups

AI Chip Vendors

A new AI workload that drives continued market adoption of GPUs and TPUs.

Relevant to

GPU & TPU makers · AI-accelerator designers · cloud-silicon programs

The Engine

Every optimizer before Questor trades completeness for time. We removed the trade.

Conventional solvers sample, approximate, and stop, occasionally returning a good-enough guess. Questor delivers the exact optimization: the complete problem resolved, exactly, and re-resolved the instant your requirements change.

Exhaustive answer, not exhaustive search

The engine returns the exact Pareto frontier—every optimal solution, provably none missing. It gets there by proving which regions cannot hold an optimal solution and discarding them whole, not by examining every candidate. No sampling, no good-enough, no hallucinations—the exact answer, every time.

Interactive by design

Change an objective and the system re-solves in place. Steer the outcome in real time, the way you’d steer a car.

Scales to large problems

Problem size changes the numbers, not the method. Runtime tracks the size of the answer rather than the size of the space—which is why a vastly larger problem can resolve faster than a smaller one.

Exact on modest hardware

A single low-power GPU is enough for exact optimization—our benchmarks run on an entry-level NVIDIA L4. More of them, or a cluster, are optional rather than required.

Optimization and refinement

A Tradeoff Map distills a full Pareto frontier to its essence, making it consumable and actionable by humans. In this case, below, from a problem space of nearly a billion billion potential regimens, optimization identifies twelve optimal regimens on the Pareto frontier, then distills this to five Paths presenting the essential tradeoffs made by the twelve optimal regimens.

Here the clinical user chose their objectives for this patient in this encounter, a senior patient experiencing falls. The Tradeoff Map informs the clinician of the best regimens with respect to these objectives — but not all objectives can be achieved simultaneously.

Tradeoff Map for the 16-medication regimen. A starting regimen scoring 10% orthostatic hypotension, 48% dizziness, 19% insomnia, 25% fatigue and 7% anemia, followed by five optimal paths: Path A reaches 0 to 1% hypotension and 0% anemia while accepting 13 to 14% dizziness; Path C accepts 7 to 8% hypotension for 1 to 3% fatigue; Path E reaches 0% fatigue.
Tradeoff Map · 16 medications · five paths from one starting regimen
The enabler

Untapped potential, unlocked.

Questor exploits the untapped potential of low-power GPUs—leveraging advanced mathematical frameworks, AI-driven inference, and insights from quantum computing to keep optimization and refinement precise, rapid, and always at the ready.

Low-power GPUs

Mathematical frameworks

AI-driven inference

Quantum insights

Two halves of one mind.

Questor is the explorative optimization to generative AI’s pattern processing— and it deploys on similar GPU infrastructure.

Generative AI

Pattern processing

Generation·Patterns·Inference
Questor

Questor

Explorative optimization

Exploration·Optimization·Distillation
Similar GPUs
Live demos

Intrigued? Sign up to get a live demo.

You know your checkboxes. Let Questor find the maximal ways of checking them.

Sign Up for a Demo
About us

About Us

Questor is an R&D lab performing basic research in advanced computation, development of next generation technologies, and is especially focused on identifying, improving and optimizing tradeoffs. The lab spins off domain-centered solutions, for example, Surveyor Health.

Co-founders

Who is behind Questor.

Erick Von Schweber

Erick Von Schweber

Co-founder

LinkedIn

From the time he was a young child Erick was fascinated by science, in particular by the universe, by human physiology and anatomy, and by machines. Over the years these evolved, with astronomy deepening to astrophysics and cosmology then to the foundations of physics, while his biological interests zeroed in on the brain, cognition and evolution, and his machine interest found expression in artificial intelligence (and also in auto engineering and racing, eventually moving from hot rodding Firebird Trans Ams to hot rodding computers).

He decided to focus primarily on mathematics to support his diverse but related interests, with secondary foci on AI and quantum physics. Disillusioned with the way AI was moving in the late 1970s and early 1980s he connected with the fellow who would become his mentor, the late quantum relativist and flagbearer of quantum logic, David Finkelstein.

Erick saw in David’s work an alternate path to understanding intelligence, cognition and consciousness, where the quantum underpinnings of mind and world merge, with applicability to real world problems.

Linda Von Schweber

Linda Von Schweber

Co-founder

LinkedIn

Linda has decades of experience in advanced technology design and management for Government and Defense contracts, where she led the coordination of eight companies and university research groups on various government programs.

As the President of Surveyor Health, she oversees project management and client delivery, ensuring seamless execution. Linda also designed the platform’s innovative interface, capable of handling up to 60 medications simultaneously for patient reviews. By organizing data in an intuitive, spreadsheet-like format, her design makes the impossible possible: reducing complex medication review time from 1 hour to just 30 minutes, significantly improving efficiency and driving cost savings for clients.

For Questor Linda collaborated with Erick and Shawn on the conceptual, theoretical and practical implication of tradeoff management and guided the UX design so that one UX can serve any domain.

Shawn Kessler

Shawn Kessler

Co-founder

LinkedIn

Shawn Kessler architected and co-developed Questor’s engine and the platform around it—from its GPU-accelerated core to the cloud infrastructure it runs on. With more than 25 years as a software architect, developer, and data scientist, he’s happiest either reasoning about system design or getting his hands dirty in the code. He holds a B.A. in Computer Science and a Master of Information and Data Science, both from U.C. Berkeley.

Shawn is also VP of Engineering at Surveyor Health where he manages new development and day-to-day operations, and is first author of the team’s peer-reviewed study in the Journal of Managed Care & Specialty Pharmacy, with co-authors from Stanford University.

All three co-founders are authors of the peer-reviewed JMCP study, alongside co-authors from Stanford University and Inland Empire Health Plan.

Working together

Four decades on the same problem.

With the backdrop of Erick’s PhD studies in quantum logic and Finkelstein’s weekly quantum topology workshop, Erick met Linda and they began working together. Before long they became a couple, and they have worked on the same problem ever since — through an early semantic network built on quantum logic rather than Boolean, a classified DARPA program, and a public prediction of where computing was headed.

When Linda’s mother nearly died due to medication side effects they decided to form Surveyor Health, applying the technology they had already spent decades building — the engine that would later be named Questor.

For the technology’s own history, see R&D Evolution.

Track record

Where the work has been proven.

First three

One of the first three mainframe downsizing projects to a PC client-server network, on Oracle v5.

DARPA

Architect on multiple programs, most classified, including SAPIENT — optimizing military networks.

1998

Coined the term “computing fabric” in a PCWeek column. Ten years later IEEE’s Computer credited them with it — vol. 41, no. 9 (PDF).

Peer-reviewed

−19.3% total cost of care in a two-year clinical study, published in JMCP.

The lab

The engine is the through-line. It was developed across four decades of research, commercialized in medication risk under the name SurveyorAI, and renamed Questor once its scope had outgrown medicine. What the clinical study measured and what Questor does today are the same engine at different stages.

Contact

Reach a person.

Get in touch
R&D Evolution

From tradeoffs to Questor.

Questor didn’t appear overnight. It is the evolution of a decades-long R&D program built on a single conviction: that tradeoffs matter—and that improving, optimizing and refining them creates substantial value.

Where things stand today

Two paths, one Big Bang.

Deploy far and wide

Questor applied across many industries through deep partnerships and licensing.

An agentic architecture

Collaborating AIs—demonstrated in Surveyor Health’s Closed-Loop Collaborative Care Intelligence: conversational AIs with people, genAIs for inference, Questor for optimization.

The “Big Bang”

Deeply integrating Questor as the explorative optimizer to generative AI’s pattern-processing.

Where did Questor come from?

It began with Information Unbound.

Questor’s roots run deep—an evolution of an R&D program launched by Linda and Erick Von Schweber. The driving idea was simple: tradeoffs matter, tradeoffs can be improved, and doing so creates real value.

You can get this—but not with that.
Nothing erudite: the everyday shape of a tradeoff. And they aren’t only human—they’re woven into the world itself, from Heisenberg’s indeterminacy to Bohr’s complementarity at the foundations of physics.
The road here

Four decades of resolving tradeoffs.

  1. Mid-1980s

    Massively parallel roots

    Early research showed how massively parallel supercomputers—the SIMD and MIMD forerunners of today’s GPU clusters, like the Connection Machine and the nCube—could be harnessed to understand, visualize and optimize tradeoffs.

    More…Less

    This involved developing a semantic network embodying Ayn Rand’s Objectivist Theory of Concepts, but with quantum logic in place of Boolean, calling it ATLAS. With an NLP front-end they discovered that users brought with them presumptions and presuppositions in stark contrast to the knowledge in the semantic network. Awareness of Scott Fahlman’s NETL system at MIT’s AI Lab illuminated the value of exceptions to inference via non-monotonic logic.

    The parallel between query lattices to handle such failures, the orthomodular lattice of closed subspaces of a Hilbert Space of quantum mechanics, and the Galois connections of concept lattices and Chu Spaces began simmering, leading eventually to Questor.

  2. Domain-independent by design

    First trials, many verticals

    From the outset the goal was a domain-independent approach—usable anywhere tradeoffs stymie progress. Early tests spanned automobile configuration, shopping, and human resources / recruitment, all showing real applicability.

  3. The consortium

    Synthesizing systems of systems

    Linda and Erick founded Synsyta LLC — “Synthesizing Systems of Systems” — a research-focused consortium coordinating eight companies and university research groups, delivering R&D and framework studies on integrated advanced computing, semantic web technologies and service-oriented architectures to DARPA, intelligence bodies and the General Services Administration.

  4. DARPA

    Battle-theater networks, in real time

    The philosophy and technology advanced onto multiple classified DARPA programs including SAPIENT with Raytheon and Telcordia—optimizing the bandwidth and availability of battle-theater wireless networks in real time.

  5. The fork

    Choosing the hardest problem

    For an open, non-classified commercial application the choice came down to drug-candidate discovery or adverse drug events—then responsible for over 200,000 US deaths a year. With Linda’s mother facing a life-threatening event, Medicare Part D launching, and licensable structured data available, they chose adverse drug events—founding Surveyor Health and branding the core technology SurveyorAI.

  6. Clinical proof

    Trade-offs matter—proven

    A two-year, peer-reviewed clinical study—run with Inland Empire Health Plan, one of the ten largest Medicaid plans in the U.S.—confirmed with high statistical significance that improving medication-risk tradeoffs cut Total Cost of Care by 19.3%, through fewer ED visits and hospitalizations. Published in the Journal of Managed Care & Specialty Pharmacy, it stands as a milestone still unmatched.

  7. Sustained validation

    Held up in the real world

    The largest federally qualified health center runs the SurveyorAI Clinical Platform daily—now in its fourth year, with a 62% reduction in ED utilization. A large Southern California hospital system saw heart-failure admits and readmits fall 45–54%.

  8. The enablers

    genAI meets the GPU era

    Two related shifts opened the broader market: generative AI as a limitless source of structured data, and the explosion of rentable GPUs. Together they let the tradeoff engine run horizontally—across any vertical—quickly.

  9. Today

    SurveyorAI becomes Questor

    To signify vastly expanded capabilities and capacity, the technology was rebranded from SurveyorAI to Questor.

Why now

The two enablers.

Generative AI as a data source

Beneath the noise about slop and hallucination is a real capability: genAI can source, structure and reconcile data into almost any computable form. The old challenge of finding usable structured data has vanished.

Find Extract Transform Structure Reconcile Elaborate

The GPU that came home

GPUs mastered the massive parallel matrix math of graphics, then of AI. A modern desktop GPU is thousands of times more powerful than the original Connection Machine. Thanks to its deep roots in massive parallelism, Questor exploits that same matrix math—a perfect pairing with generative AI.

From a proof point in the hardest vertical to a domain-independent engine.

Surveyor Health proved the premise. Questor carries it everywhere—explorative optimization for generative AI’s pattern-processing, on similar GPU infrastructure.

See the applications
FAQ

Questions, answered.

What is Pareto Optimization?

Vilfredo Pareto, an engineer and economist, looked into systems, like economies, that involved tradeoffs on characteristics, properties, relationships and measurements.

He attended to questions such as what constitutes a fair, equitable distribution of wealth in society, and thought deeply about wealth distributions where in order to make one individual better off someone else would have to be made worse off. In the case he pondered over, one characteristic would tradeoff against another. His thinking developed into an identification of optimal configurations like these, optimal in the face of conflicting objectives.

This was eventually generalized to the idea of identifying all optimal options/configurations such that none can be improved in one respect without causing a worsening in another. More formally, a state is Pareto optimal if no feasible alternative exists that is strictly better for at least one objective and at least as good for all others. The set of all Pareto Optimal solutions is called the Pareto frontier or the Pareto front.

Pareto’s insights have been embodied under numerous labels, including Multi-Objective Optimization (MOO), Many Objective Optimization (MAOO) and Multi-Criteria Decision Making.

How can it be that fast?

Because it never examines most of the possibilities — it doesn’t need to. The work grows with the size of the answer, not the size of the problem.

Compare two runs. A 16-medication regimen is drawn from 7.2×1017 possible combinations and resolves in 0.003 seconds, because only 12 regimens are optimal. A travel itinerary is drawn from 9.95 million combinations — a search space 72 billion times smaller — and takes 0.655 seconds, because 684 itineraries are optimal. The vastly larger problem finished more than 200 times faster. Runtime follows the answer, not the question.

Both figures move with the problem as you set it up. Restricting which alternatives the optimizer may consider shrinks the search space, the frontier and the runtime together — the numbers above are single runs with every alternative allowed, not fixed specifications.

This is why dividing a combination count by a runtime produces a meaningless figure. No classical supercomputer evaluates 1020 candidates per second, and Questor does not try to. It proves which regions of the space cannot contain an optimal solution and discards them whole, so only a vanishing fraction is ever examined — and the frontier that comes back is still complete.

What is the difference between exact and approximate Pareto optimization?

Exact optimization returns every optimal solution on the Pareto frontier, provably. Approximate methods explore only a fraction of the possibilities and hope the best ones are among them.

Finding the full frontier is a computation of enormous scale when attempted via brute force. Optimizing a system of 16 parts, each chosen from 16 components, means considering all 1616 possible systems — roughly 2×1019, about 20 billion billion. Worse, identifying which are Pareto optimal across, say, four objectives (minimize cost, time, unavailability, weight) requires comparing each against the others — about 3.4×1038 comparisons, far beyond the reach of the world’s fastest supercomputer.

Brute force is exact but infeasible. Heuristic pruning helps, but because it discards regions that may still hold optimal solutions (often referred to as concave solutions not on the hull of the convex set), what comes back is an approximation — and past toy problems it remains far too slow for interactive use. This is the path generative AI takes.

Because the results are so valuable, many approximate methods have been developed — evolutionary and genetic algorithms, particle-swarm methods, simulated annealing, and others. Rather than explore the entire space, they effectively sample a small part of it, trading exactness for speed.

Questor eliminates rather than samples. It discards only those regions it can prove cannot contain an optimal solution, so nothing optimal is ever lost and the frontier that returns is complete. This is why Questor’s runtime tracks the number of optimal solutions rather than the size of the search space.

Furthermore, for optimization problems that are larger than toy demonstrations, there exist way too many optimal solutions, creating an infoglut of options that are far too numerous for human consumption. But Questor is more than exact optimization, it’s also refinement, distilling the frontier into a tradeoff map fit to inform human, and machine, decision making.

What is a Path in a Tradeoff map?

A Path in the Tradeoff map is a group of optimal solutions that make broadly the same tradeoffs, so you can choose between a handful of directions rather than thousands of individual answers.

The Tradeoff map refines the exact Pareto frontier via an algebraic distillation process.

In the 16-med example the engine returns five paths. One drives orthostatic hypotension down to 0–1% while accepting 13–14% dizziness; another accepts 7–8% hypotension in exchange for better fatigue. A single path may hold more than one optimal option set. Change the objectives and you get different paths — the map reflects what you asked for.

Choosing between paths is a judgement, not a computation. Every path is optimal and none is strictly better than another — that is what being on the frontier means. Narrowing from a Path to a specific system/regimen is what the Valid Choice Navigator is for, providing the user with the opportunity to fully specify an optimal system.

Why do some results look different?

Because every figure shown here comes from a single run, and a run reflects the objectives chosen for it.

Change which objectives are included, or restrict which alternatives the optimizer may consider, and the search space, the frontier and the runtime all move together. A different set of objectives yields a different set of paths — that is what the tool is for, not a defect in it.

So the numbers here are illustrations of the method rather than fixed specifications. Two captures of the same demo, taken with different objectives, will disagree with each other and both will be right.

Why hasn’t this been done before?

Because the field of computational optimization concluded exact optimization at this scale was impossible, and spent decades improving the next best thing instead.

Furthermore, the team behind Questor developed its methods for massively parallel operation from the beginning.

Questor came at this from a different but valid direction: treating the frontier as something to be constructively derived rather than discovered, so completeness is a property of the method instead of a by-product of looking everywhere.

How is Questor different from generative AIs?

Generative AI is mathematically formal in how it computes, but not in what it concludes — which is why it hallucinates. Questor is formal in both.

Generative AIs and LLMs evolved from artificial neural networks, described as layers of parameters — weights and biases — trained (via back-propagation, gradient descent, and attention) to detect patterns in their inputs and respond to prompts. Their internal calculations are mathematically sound, but their pattern-based outputs are not formal in the mathematical sense, which is the root of hallucinations and “AI slop.”

Questor is mathematically formal in both its computation and its results. Where generative AI is general-purpose across the prompts it can handle, Questor is special-purpose: exact multi-objective combinatorial optimization. Ask a generative model to perform an exact optimization and it will plan, write, and run code to attempt it — but that generated code lacks Questor’s innovations for delivering exact, human-scale results interactively. Questor delivers them.

How is Questor a general-purpose optimizer?

Questor is a special-purpose engine for optimization problems that applies across virtually any domain, because the setup always takes the same shape.

Questor takes a system configuration specification — the parts of a system and the candidate components that can fill each part — along with data on those components (their properties, attributes, features, and interactions) and the functions that evaluate a system-level attribute from its components’ attributes. Every industry and vertical has its idiosyncrasies, but we have found these inputs fall into similar groups in terms of data setup.

How does Questor use generative AI?

One or more generative AIs can perform the specific work for each vertical application: searching for, finding, extracting, reconciling, normalizing, and formatting the data that drives Questor.

This is optional for each application, in that Questor can be pointed to or integrated with structured data set resources directly. This can be done in real-time to pull in the most recent information such as the current menu options from a restaurant. It can also be done at design time or periodically, then monitored live for changes and updated.

What does this have to do with quantum computing?

Erick, our co-founder, was mentored by the quantum relativist David Finkelstein, who not only gave us the modern conception of astronomical black holes (he called them “unidirectional membranes”) but also a very deep understanding of the physics that underlie logic itself.

David identified that some classical algorithms, operating on standard, classical computers, exhibit the telltale signs of quantum: superposition, entanglement and complementarity. We identified one class of such algorithms and have employed David’s insights to accelerate them.

What is a GPU?

A GPU, a Graphics Processing Unit (i.e., chip), is a processor built to perform large matrix-math operations quickly and in parallel.

The “graphics” in GPU comes from its original purpose: the large matrix operations behind 3D graphics — lighting, viewpoint and perspective transformations, and scene rendering. It turns out that training and running LLMs and generative AI require the same kinds of matrix operations, at even larger scale — which is why GPUs power modern AI.

We use the term “GPU” broadly to include related matrix-acceleration processors such as TPUs (Tensor Processing Units) and AIPUs (AI Processing Units).

What does QUESTOR stand for?

Quantum Exact System Tradeoff Optimization and Refinement.

We’ll start with “Exact” then return to “Quantum”.

Exact because the engine returns the complete Pareto frontier rather than a sample of it.

System because what gets optimized is a configuration — parts, and the candidate components that fill them.

Tradeoff because Questor performs its optimization in the face of potentially many conflicting objectives.

Optimization is the process of making a system process or design as effective, efficient or functional as possible. It means getting the best possible result or highest performance by changing certain variables while working within fixed limits or rules.

Refinement because the frontier is not the end of the work: it’s distilled to its essential tradeoffs.

Quantum refers to quantum principles that appear inside certain classical algorithms, in this case, tradeoff optimization algorithms — not to quantum hardware. Questor runs on ordinary GPUs.

Partnerships & Licensing

Bring Questor to your industry.

Questor is domain-independent by design, and there are two ways in: partner with us to co-develop an optimization-powered product for your domain, or license the engine to embed and run it yourself—on the low-power infrastructure you already have.

Partner with us See it live

Deep partnerships

Co-develop domain-specific solutions with our team, from first model to production.

License the engine

Embed Questor directly into your platform and put exact optimization behind your own front door.

Your infrastructure

Runs on entry-level GPUs, on-prem or in your cloud. Clusters optional, no lock-in.

Live demos

Intrigued? Sign up to get a live demo.

You know your checkboxes. Let Questor find the maximal ways of checking them.

Sign Up for a Demo
The science

Where the method comes from.

The science

Pareto optimization

Vilfredo Pareto, 1848 to 1923
Vilfredo Pareto
1848–1923

When several objectives pull at once and no single solution is ideal in every respect, Pareto optimization identifies a frontier: all and only the optimal solutions. On that frontier, any attempt to improve a solution in one way necessarily makes it worse in another—the everyday shape of a tradeoff. You can get this, but not with that. Optimizing a system against conflicting objectives is known variously as multi-objective optimization, multi-objective combinatorial optimization, many-objective optimization, or Pareto optimization. The name honors Vilfredo Pareto, the Italian engineer and economist.

Exploits quantum principles that appear inside certain classical algorithms.
Computational Complementarity can exploit these quantum properties to speed processing of the Pareto frontier.

Quantum inspiration

Erick Von Schweber with David Finkelstein outside the School of Physics at Georgia Tech, December 2011.
Erick with David Finkelstein
1929–2016
Georgia Tech, Dec 2011

This line of work follows the late David Finkelstein, the quantum relativist who mentored Erick, and who identified that certain classical algorithms running on ordinary computers show the telltale signs of quantum. Our research has revealed that a Pareto frontier is itself a kind of quantum system, exhibiting superposition, entanglement, and complementarity. Guided by Computational Complementarity—and by what quantum computing has taught us about isolating and carefully controlling interactions—computations on and around the frontier can be accelerated.

Quantum computing has revealed a key principle—for a quantum system to perform computation, its superposition of states must be isolated from interactions that may untowardly and indeterminately alter it, or even cause it to collapse before it’s performed useful work. The two-slit diffraction experiment illustrates this. Any attempt to measure which slit the quantum passes through collapses the superposition and eradicates the interference pattern that is a telltale of quantum.

This is one of the ways Questor resolves the frontier in seconds—the acceleration long promised by quantum AI and quantum optimization, delivered on classical, low-power GPUs.

No quantum computing hardware needed
Closed-loop care

Closed-Loop Collaborative Care Intelligence

Beyond using AI to summarize the past, this platform integrates conversational agents, Bayesian risk modeling and exact optimization to co-engineer the patient’s future.

In a peer‑reviewed clinical study, published in the Journal of Managed Care and Specialty Pharmacy, pharmacists using the Surveyor Health Clinical Platform, powered by SurveyorAI, were able to weigh medication risk tradeoffs at‑a‑glance and act on them quickly, for targeted high‑risk patients. After 14 months and nearly 7,500 patient encounters the result was a statistically significant reduction in the Total Cost of Care of 19.3%. Also observed was a doubling of pharmacist productivity. Were this technology and workflow deployed for all 5 million Americans at similar risk, the savings would amount to over $32 billion annually.

The fly in the ointment is the limited availability of human clinicians. Even with SurveyorAI’s doubled productivity, a national‑scale rollout would require about 5,000 patient engagement assistants plus another 5,000 clinical pharmacists — clinicians that are just not available.

The solution is to integrate a patient engagement AI agent to conduct patient intake and medication reconciliation — but not to dispense medical advice — with the Questor optimization engine to determine optimal medication options that providers can choose from for adoption. We call this Closed‑Loop Collaborative Care Intelligence.

Population surveillance
Patient intake & MedRec
Patient risk assessment
Symptoms confirmed with patient
Therapy refinement
Patient–provider shared decisions
SurveyorAI
SurveyorAI
PharmD
PharmD
SurveyorAI
SurveyorAI
PharmD
PharmD
SurveyorAI
SurveyorAI
PharmD
PharmD
SurveyorAI
SurveyorAI
Aimi
Aimi
Questor
Questor
Aimi
Aimi
SurveyorAI
SurveyorAI
Aimi
Aimi

19.3% reduction in total cost of care, achieved with clinical pharmacists on SurveyorAI. Now being made autonomous.

Bayesian models assess medication risk across the whole population, and target the highest-risk patients.

Aimi converses with the patient to reconcile their medications. It cannot hallucinate, and never gives medical advice.Sponsored by Amazon AWS · Developed with Caylent

The risk assessment updates, then generative AI correlates it with diagnoses and lab history to find the material risks.

Aimi confirms with the patient whether those risks are actually manifesting.

Exact optimization sorts conflicting goals and manifold risks into the key tradeoffs a provider can act on.

Aimi carries the decision between patient and provider, then monitors progress.

Aimi and Questor deliver it autonomously — at a scale the pharmacist workforce cannot reach.

The JMCP study results: 19% lower total cost of care and $6.4M annual savings for every 1,000 high-risk patients targeted. Claims data and EHR data across a population of patients. Aimi in conversation with a patient, going through their medications one by one. A matrix of interactions between the patient's medications. Aimi asking the patient whether they have experienced dizziness or lightheadedness. The tradeoff table for the patient's regimen, showing the starting regimen and several optimal paths. Aimi relaying that Path A eliminates the risk of orthostatic hypotension, and the provider adopting it.
Step 1 of 8

Proven outcomes

IEHP, one of the ten largest Medicaid plans in the U.S., supported by clinical pharmacists and medical assistants, achieved a 19.3% reduction in total cost of care using this process and SurveyorAI. Now, with additional AI and optimization, we are making it autonomous.

The JMCP study results: 13% fewer drug duplications, 15% fewer drug-drug interactions, 9% fewer hospital admissions, 10% fewer bed days, 15% fewer ED visits, 17% lower medication cost and 19% lower total cost of care, with $6.4M annual savings for every 1,000 high-risk patients targeted.
Step 2 of 8

Population surveillance

SurveyorAI’s probabilistic (Bayesian) models continually assess medication risk across the whole population, and target high-risk patients for deeper analysis.

Claims data and EHR data across a population of patients.
Step 3 of 8

Patient intake & medication reconciliation

Aimi, an AI agent, converses with patients by text and voice, performing intake and medication reconciliation. Aimi cannot hallucinate and never provides medical advice.

Sponsored by Amazon AWS · Developed with Caylent

Aimi in conversation with a patient, going through their medications one by one.
Step 4 of 8

Patient risk assessment

The Bayesian risk assessment updates to reflect what reconciliation found. Generative AI then correlates those medications with diagnoses and lab history to determine which risks are material.

A matrix of interactions between the patient's medications.
Step 5 of 8

Symptoms confirmed with the patient

Informed by the updated analysis, Aimi confirms with the patient whether those material risks are actually manifesting — by text, or by voice.

Aimi asking the patient whether they have experienced dizziness, lightheadedness, diarrhea or nausea.
Step 6 of 8

Therapy refinement

Exact optimization is applied to the patient’s regimen, sorting out conflicting treatment goals and manifold risks, and distilling the key tradeoffs so the provider can decide.

The tradeoff table for the patient's regimen, showing the starting regimen and several optimal paths.
Step 7 of 8

Patient–provider shared decisions

Aimi facilitates the collaboration between patient and provider on goals and therapies, then monitors progress.

Aimi relaying that Path A eliminates the risk of orthostatic hypotension and anemia, and the provider adopting it.
Step 8 of 8

Autonomous delivery

Aimi and Questor together make these outcomes and savings available autonomously — at a scale the pharmacist workforce cannot reach.

The loop closes, and repeats

Continuous, autonomous care intelligence

The key components needed to enable Closed-Loop Collaborative Care Intelligence have been developed, principally the Aimi agent to converse with patients and providers, the SurveyorAI Clinical Platform and the Questor optimization engine. Surveyor Health is integrating these facilities, seeking beta partners to guide deployment of the integrated system, operational testing and publication.