The problem is technically ambiguous.
The outcome matters, but the architecture, algorithm, research direction — or even the right framing — is unclear.
Independent Applied AI / R&D Consultant
For technical problems that don’t fit neatly into a job description.
I help technical teams, founders, R&D groups, engineering leaders and investors resolve difficult AI and R&D questions through bounded investigations, validation work, technical advisory and fractional direction.
Free and focused on fit and a practical next step.
A brief introduction01 / Perspective
Click the picture to play. Full screen and volume are available in the controls.
How I approach difficult technical problems1:52
Open the dedicated video page and transcript →
Read the video transcript +
Some valuable problems arrive without a clear playbook. I’m Brian Theory, an independent Applied AI and R&D consultant, helping ambitious teams turn unusual ideas into useful systems.
Through August 2025, I led software engineering work supporting NASA Ames Thermophysics Research. Before that, I worked in computational neuroscience and contributed to several publications, including a 2024 study I led using explainable machine learning to develop shorter psychometric measures.
I developed correspondence matrices, an original structural approach to Boolean computation, and actively maintain its public GitHub repository. It includes the implementation, reproducible comparisons with established algorithms, and detailed explorations of potential applications, from formal verification and rule systems to symbolic AI. If you’re technically curious, you can examine the code and evidence directly.
Codex and Claude Code are part of my development process, helping me explore alternatives and implement faster. I frame the problem, choose the approach, connect the pieces, and test whether the system behaves as intended.
That approach applies to multi-agent automation, evidence-grounded research tools, and scientific or creative software, especially where requirements evolve as the work reveals new information.
I take on focused consulting and fractional R&D engagements for technically difficult work. I’m drawn to learning new domains, investigating open problems, and adapting as research moves forward.
If you’re building something unusual or unproven, visit CorrespondenceMatrices.com and tell me what you’re working on.
Current website: CorrespondenceMatrices.com. Implementation, papers and limitations are linked in Selected Technical Work and the CV.
Four proof anchors
near the work
Correspondence Matrices →Public implementation, reproducible
benchmarks and correctness checks
benchmarks and correctness checks
Nexus Agent Operations →Human approval, durable execution,
provider evidence and auditability
provider evidence and auditability
Local Multi-Agent AI →Air-gapped models · hybrid retrieval
configurable agent workflows
configurable agent workflows
Paper Workbench →Public implementation · provenance
audits and human review
audits and human review
01 / The right kind of help
When the usual playbook
isn’t enough.
Start with a useful way into the problem,
grounded in the decision that matters.
You need feasibility before commitment.
Find out which assumptions hold before committing substantial engineering time, capital or organizational attention.
Your team needs another level of thinking.
Bring additional technical depth into an important investigation, alongside the people who already know your systems.
The problem crosses disciplines.
Work that falls between software, AI, mathematical modeling, cognition and systems rarely fits a conventional brief.
An idea needs to become an experiment.
Turn an unfamiliar concept into a model, architecture, prototype or technical question that can actually be tested.
You’re stuck.
Get an independent technical perspective. Revisit the assumptions, change the representation and identify another way through.
02 / Fractional R&D, explained
Technical capability.
Without a permanent role to fill.
Research and technical capability when you need it, without creating a permanent role before the problem justifies one.
Explore fractional direction ↗An investigation with a decision at the end.
Several intensive sessions, a feasibility study, an independent technical review or due diligence. Start with the uncertainty that matters most.
Test the uncertainty, not the entire product.
When the question is ready, use a model, technical spike or limited proof of concept to generate evidence for the next decision.
Embedded thinking, as the work develops.
Fractional R&D direction, collaboration with your existing team, or temporary guidance of an exploratory initiative.
Start with one decision. A hands-on validation sprint is a selective follow-on, limited to one active engagement at a time.
Ways to work together
Start small enough to decide.
Available starting point
Feasibility & Decision Sprint
Resolve one consequential uncertainty before committing to a build.
$9,000 standard sprint · USDSelected follow-on
Technical Validation Sprint
Run the smallest experiment that can change the decision.
From $20,000 · selected follow-on · USDBroader technical work
Extended Technical Investigation
Investigate a wider technical programme without forcing it into a standard sprint.
Custom scope · USDReserved capacity
Fractional Applied AI / R&D Direction
Technical direction and continuity for work already in motion.
From $8,000/month · USD03 / Capabilities
From an open question
to a working answer.
An engagement can move through several modes.
The question determines what the work needs.
Explore
Find the question worth answering.
- Technical feasibility & due diligence
- Research synthesis & investigation
- Applied ML evaluation
- Mathematical & algorithmic R&D
Design
Give the problem a useful structure.
- AI systems & software architecture
- Local, private & on-prem AI
- Retrieval & knowledge systems
- Algorithms & computational models
Build & test
Replace assumptions with evidence.
- LLM systems & agentic workflows
- Bounded proofs of concept
- Simulations & experiments
- Evaluation & technical validation
Advise & guide
Help the team make its next move.
- Fractional Applied AI / R&D direction
- Research-to-production planning
- Independent technical review
- Embedded team collaboration
Explore AI feasibility, technical validation and fractional R&D services →
The cross-disciplinary advantage
More ways to represent
and attack a problem.
Some problems are difficult precisely because they sit between disciplines. Computer science, AI, computational neuroscience, mathematics and systems thinking offer different models of the same underlying challenge.
The value is knowing when to borrow an abstraction, question a familiar approach, or connect research to an implementation.
The point isn’t breadth for its own sake.
It’s having more ways into the problem.
ComputationMathematicsCognitionSystems
One shared focusThe problem that matters
Models & experiments → Evidence
Different perspectives. A concrete next step.
04 / Selected R&D and systems work
The thinking, made concrete.
Original algorithmic research leads into governed and local AI, evidence-controlled research, published methods and engineering systems.
01 / Algorithms & structural reasoningCorrespondence MatricesA different representation of Boolean computation—developed into a public implementation, reproducible comparisons and independent correctness checks.
situation
Investigating alternative ways to represent and reason about Boolean computation.
thinking
The work explores a structural representation through correspondence matrices, with potential applications including formal verification, rule systems and symbolic AI.
work
Developed the representation, a Python implementation and reproducible benchmark tooling that compares specific operations with SymPy, ROBDD, Espresso and other backends. Correctness checks run outside the timing windows.
result
An implemented approach with source, installation notes, benchmark commands and generated reports that can be examined and rerun. No general comparative performance advantage is claimed.
Scope & limitations
Correctness is checked against independent truth-table evaluation. Benchmark timings compare specific implementations and operations; they are not a claim of a general speed advantage.
02 / AI operations · Exact action authorizationNexus Agent OperationsNexus — a human-governed operational platform for real provider workflows with typed actions, approval-bound authority and auditable outcomes.
situation
Agent systems become operationally risky when proposing an action, approving it and executing it are allowed to blur into one model-controlled step.
thinking
Keep research, planning and drafting useful while making each state-changing operation pass through typed validation, identity checks, human approval and an exact dispatch permit.
work
Built Python and FastAPI services, a Next.js operator console, typed operation contracts, durable Neo4j queues, approval gates, audit trails, pacing, suppression and emergency-stop controls around real provider workflows.
result
Separated planning, review, dispatch authority, provider outcome and audit evidence into distinct system states across live email, inbox and bounded integration workflows.
Scope & limitations
The public account documents the operational architecture and verified workflows while keeping the private implementation and provider credentials out of the portfolio.
03 / Applied AI · Engineering automationLocal Multi-Agent AIEngineering Documentation & Analysis — configurable agents, local models, hybrid retrieval, deterministic NLP and engineering-specific workflows.
situation
Large collections of requirements, component documentation, procedures and vendor material needed evidence-based analysis in an environment where unrestricted cloud-AI processing was inappropriate.
thinking
Build the system around locally hosted models and technical evidence, with retrieval and conventional NLP carrying structure that should not depend entirely on an LLM.
work
Built a YAML-configured Python agent framework using Ollama, FAISS and SentenceTransformers, BM25 hybrid search, spaCy and NLTK extraction, persistent indices, caching, asynchronous processing and engineering-specific analysis.
result
Produced a local engineering-document environment for search, extraction, requirements organization, cross-document synthesis and specialized agent workflows without making cloud inference a prerequisite.
Project scope
The framework ran in an air-gapped engineering environment, with local models, retrieval indices, deterministic NLP and configurable workflows operating on locally available technical sources.
04 / Evidence-controlled research · Applied AIPaper WorkbenchA public research environment connecting sources, claims, human decisions, audits and exports through explicit provenance.
situation
AI-assisted research can make drafting faster while making it harder to see where a statement came from, which evidence supports it and what a person actually approved.
thinking
Treat provenance, support state and human decisions as part of the research object rather than metadata added after the work is written.
work
Conceived, architected and built a local-first research workbench with grounded dialogue, claim-to-evidence links, manuscript workflows, approval-bound AI actions, deterministic audits and provenance-bearing exports.
result
Produced a public implementation with 101 offline tests, capability and evaluation reports, and an inspectable chain from source material through claims and decisions to exported research artifacts.
Scope & limitations
The current system is a local, single-machine research environment. Its published capability matrix distinguishes implemented, partial, planned and out-of-scope functionality.
05 / Explainable ML & measurementExplainable ML for Psychometric MeasuresA 2024 research study applying explainable machine learning to psychometric measurement.
situation
Investigating shorter psychometric measures, at the intersection of computational methods and behavioral research.
thinking
Using explainable machine learning to examine measurement, rather than treating a predictive model as an uninterpreted endpoint.
work
First-authored a 2024 open-access study using neural networks and Shapley values to guide item retention for abbreviated versions of the Penn State Worry Questionnaire in a sample of 3,906 adolescents.
result
Items with higher SHAP importance produced stronger abbreviated configurations, and item importance changed with the portion of the worry distribution being targeted. The study compared the method with a factor-analytic five-item questionnaire using the same dataset.
Scope & limitations
The study used one adolescent sample without independent external validation. The paper also notes that SHAP interpretations can be sensitive to correlated features.
06 / Systems integration · Industrial data acquisitionPROFINET Data Path EngineeringPC Acquisition Host ↔ Siemens S7-1500 — integration across PCIe hardware, embedded C, process-data mapping and PLC configuration.
situation
A NASA Ames arc-jet acquisition upgrade needed a PC host to exchange process data with a Siemens S7-1500, but vendor demonstrations did not represent the application data model and device-side definitions did not initially match the PLC configuration.
thinking
Treat the interface as one data path whose firmware definitions, GSD/GSDML description, host application and PLC configuration must all represent the same process-data map.
work
Defined two separately managed 128-element FLOAT32 regions with sequence counters, reworked the host control loop, aligned GSD/GSDML modules with the compiled ADI map, and added simulated-sensor, timing and documentation support.
result
Established a working and documented PROFINET path and delivered the application-specific software, interface definition, simulation support and integration documentation for incorporation into the acquisition system.
Project scope
The project covered the complete application path from device and firmware setup through process-data mapping, PLC alignment, simulation, timing instrumentation and integration documentation.
07 / Scientific computing · Applied mathematicsQuasi-1D Thermophysics ModelingArc-Jet Calorimetry — translation of a published conical-frustum inverse heat-conduction model into LabVIEW.
situation
Conventional planar finite-slab inverse analysis does not account for the geometric and lateral-conduction effects introduced by curved calorimeter bodies.
thinking
Translate Driver, Philippidis and Terrazas-Salinas’s published quasi-1D formulation into the existing calculation environment while preserving its mathematical structure and numerical boundary conditions.
work
Worked through the spherical-coordinate heat equation, geometric terms, A–D coefficients, tridiagonal system, boundary conditions and software choices, then integrated the formulation into LabVIEW.
result
Produced a LabVIEW implementation providing a path for evaluating curved-body heat conduction rather than relying solely on the planar finite-slab approximation.
Project scope
The project translated Driver, Philippidis and Terrazas-Salinas’s published quasi-1D formulation into the existing LabVIEW calculation environment for geometry-aware calorimeter analysis.
These research and engineering examples document the question, Brian’s contribution, the implementation, the available evidence and the boundaries of each conclusion.
05 / How I work
Enough structure to make progress.
Enough flexibility to learn.
Understand
Define the real technical question, the constraints and the decisions that matter.
A shared problem frameInvestigate
Analyze the system, challenge assumptions and identify promising paths.
Evidence and alternativesMake it concrete
Develop the architecture, models, experiments or prototypes the question needs.
Something the team can useTransfer or continue
Hand over the work clearly, or stay involved fractionally as it develops.
A clear next step↳ Uncertain problems can enter before the final scope is known. The first task may be finding the right question.
06 / About Brian
A technical generalist
by background.
A problem specialist
by practice.
I work where the question doesn’t yet come with a method.
I’m Brian Theory, published academically as Brian Droncheff. My background connects software engineering, computational neuroscience and original work in Boolean computation. From 2022-12-11 through 2025-09-15, I worked as Software Engineering Lead supporting NASA Ames (Via Amentum) Thermophysics Research.
My longer-running B-Theory research programme explores mathematical representations of self-awareness. Correspondence Matrices originated within that research before becoming a separate computational project.
That breadth shapes how I work: decompose the unfamiliar, look for a useful representation, test the important assumptions, and move between research and implementation.
I’m interested in technically difficult work with people who are willing to investigate it carefully. Sometimes the useful contribution is a prototype. Sometimes it’s a different question.
A useful addition to your side of the table
Founders & CEOs
Evaluate an ambitious idea, understand technical risk and find a credible path toward implementation.
CTOs & engineering leaders
Add technical bandwidth for investigations and architecture outside the team’s normal operating envelope.
R&D & innovation teams
Explore unconventional directions and connect research to experiments and implementation.
Investors & boards
Get independent analysis of feasibility, technical assumptions and the risks behind a proposal.
07 / Notes on the work
A window into the thinking.
Notes on framing, experiments
and making technical decisions.
Problem framing
A feasibility question is a decision in disguise.
Why “can we build it?” needs a constraint, a decision and a useful stopping point.
Read the note ↗Experimental R&DPrototype the uncertainty, not the entire product.
The most useful prototype is the smallest experiment that can expose a consequential failure.
Read the note ↗20-minute Fit & Scoping Call · Free
Have a difficult
technical problem?
A short conversation to understand your goals and constraints, establish fit, and identify a useful next step. Technical design begins within an agreed engagement.
Choose an available time for a fit call, or send a project inquiry below.
Calls are confirmed through Google Calendar.