01
Reliable Reasoning
Evidence-backed answers and adversarial evaluation: can a system's conclusions survive contradiction, poisoned context, and provenance traps.
Truth Computing
Research
Applied AI research and engineering: how AI systems reason, retrieve evidence, and act reliably in environments where mistakes are expensive. Three directions, and the projects built from them.
Status terms: Research — exploring a question, no committed system yet. R&D — building and testing a prototype for internal validation only. Prototype — a working system, not yet exposed to real users at volume. Production hardening — a real system carrying real work for a specific design partner, being hardened before wider exposure. Deployed — live and reachable at a public or partner-facing URL today.
Research Directions
Three directions, one question underneath.
Can an AI system be trusted with a consequential decision. Each direction attacks a different part of that question.
02
Agentic Systems
Systems that plan, act with tools, and review their own work, with a human gate on every consequential action.
03
Efficient Intelligence
Training and steering models to spend compute where it changes the answer: curriculum learning, reinforcement learning, and fine tuning past the limits of supervised training.
Applied Research
A reasoning and verification engine.
An end to end engine that takes a PRD and a customer transcript and carries them to an evidence backed, executed, and validated outcome.
Give the engine the problem and the context around it. It works out what matters, turns the goal into a plan, carries the plan out with the right tools, and reviews its own work. Answers arrive with their supporting evidence attached. Consequential actions stop at a human.
- Evidence in, evidence out. Every conclusion carries its sources. That is the architecture, not a setting.
- Lean by architecture. Compute goes where it changes the answer and nowhere else. Cost is a constraint on the design, not a pass we make afterward.
- The foundations are technical. Founded by an engineer with a Stanford computer science background in AI and prior experience working on large scale data systems for language models.
- It compounds. The engine folds in new AI research and the lessons of its own completed work. Every finished run informs the next one.
Capabilities
What we can help with.
Six areas. When a problem comes up, this is what we bring to it.
01
Research & Validation
Designing experiments, stress testing claims, and separating a true answer from a plausible one.
02
Applied ML
Taking a model from notebook to product: data pipelines, fine tuning, retrieval, and inference.
03
Systems & Infrastructure
Building the substrate that keeps things fast, observable, and correct under load.
04
Algorithms & Theory
Reaching for the right abstraction: the proof, the bound, the structure that makes a hard problem tractable.
Projects
The work behind each direction.
Research, systems, and applied ML built from real problems. Each project is a head start on the next one.
Reliable Reasoning
-
Adversarial Legal / Crypto Benchmark Suite
A 35-item blind evaluation set for legal and cryptocurrency retrieval, contradiction surfacing, provenance traps, poisoned context, self-validation, and problem prioritization. Source provenance is visible in the corpus manifest: court opinions plus SEC, OFAC, IRS, FinCEN, FATF, MiCA, EIP, and Bitcoin sources.
TC-BENCH-2026-001 · Truth Computing · corpus verified 2026-06-17
R&D -
Colossus Local RAG Pipeline Prototype
An R&D methodology prototype for local and mock data: ingestion, retrieval, grounding checks, and evaluation loops used to study how knowledge systems fail before any production claim is made.
TC-ENG-2026-001 · Truth Computing · repo record last updated 2026-08-20
R&D -
Benchmarking Medical AI for Strabismus Detection
Establishes a reproducible machine learning baseline for detecting strabismus from clinical data, giving ophthalmology research a measurable benchmark to build on. This is research code only; it is not a medical device, has not been cleared or approved by the FDA or any regulator, and is not intended for clinical use, diagnosis, or treatment.
TC-RN-2025-001 · Matthew Torre · published 2025-02-25 (GitHub repository creation date)
Published GitHub ›
Agentic Systems
-
Truth Computing Health: Clinical AI for Vision Care
Prototype work for a vision-care design partner: human-gated AI for intake, imaging workflows, and clinical documentation patterns, each reviewed by a clinician before it enters the record.
TC-RN-2026-001 · Truth Computing · date not published
Prototype -
Feynman: Democratizing Elite AI Education
Rebuilds university-level AI coursework as a Learning Ladder, the same concept met at five heights from a kindergarten spark to the real thing, so every student finds a rung they can stand on. Free for first-generation and low-income students, rolling out first in Jurupa Valley and Montclair.
TC-ENG-2026-002 · Truth Computing · repo record last updated 2026-07-13
Launching Announcement › -
Swish: Production Backend at Consumer Scale
The server-side backbone powering a live consumer application, engineered for reliability from day one. Spans API design, data modeling, and service architecture in production Python.
Founder engineering (personal/contract) · Matthew Torre · repo created 2025-12-13, last pushed 2026-01-03
Active GitHub ›
Efficient Intelligence
-
Elo-Driven Curriculum RL for Reasoning Models
Trains reasoning models faster by serving problems matched to their live skill level instead of random difficulty. Extends RLOO with an Elo rating system and a Gaussian scheduler that ramps difficulty as the model improves, with full Elo metadata logging for reproducibility. Outperforms SFT and vanilla RLOO baselines on pass@k (Stanford CS research with Donnie Raymond).
Founder research (personal) · Matthew Torre · repo created 2026-04-28, last pushed 2026-06-08
Research GitHub › -
Steering LLM Behavior Beyond Supervised Limits
Shapes language model behavior where supervised fine tuning stalls, using reinforcement learning to align outputs with human intent. Combines reward modeling, policy optimization, and preference alignment into a repeatable recipe for more reliable, controllable AI systems.
Founder research (personal) · Matthew Torre · repo created 2026-04-28, last pushed 2026-06-08
Active GitHub › -
Optimizing Decisions Under Uncertainty with MDPs & POMDPs
Tackles sequential decision making where outcomes are uncertain and information is incomplete, the core challenge of real-world autonomy and operations. Benchmarks MDPs, POMDPs, and policy search head-to-head on a structured problem to isolate which framework delivers the best solution quality (Stanford CS238 capstone).
Founder research (personal/academic) · Matthew Torre · repo created 2025-11-11, last pushed 2025-12-04
Research GitHub › -
Probing Quantum Advantage on NP-Hard Optimization
Applies the Quantum Approximate Optimization Algorithm (QAOA) to the NP-hard Traveling Salesman Problem, the canonical benchmark for real-world routing and logistics. Benchmarks quantum approaches against classical solvers to pinpoint exactly where quantum advantage emerges at scale.
Founder research (personal) · Matthew Torre · published 2024-06-24 (GitHub repository creation date)
Published GitHub ›
Applied & archival work
-
Demystifyd: Launching a Global Career Platform
Connects students and professionals across borders as they build their careers, taken from concept to live product. Drove discovery, product strategy, and go-to-market for the web platform.
Founder work (personal/contract) · Matthew Torre · date not published
Deployed Visit › -
EZRecruit: Automating College Athletic Recruiting
Gives university varsity coaches a single proactive hub to consolidate recruit data, automate updates, and track every interaction. Built end-to-end from discovery through product strategy, design, and financial modeling (Stanford CEE250).
Founder work (personal/academic) · Matthew Torre · date not published
Product Product Review › -
Data-Driven Player Valuation: Selecting a World XI
Turns raw player performance data across global leagues into principled valuation, selecting an optimal World XI. An end-to-end pipeline spanning preprocessing, model training, and predictive analytics, with logistic regression driving transparent, defensible selections.
Founder research (personal) · Matthew Torre · published 2024-06-24 (GitHub repository creation date)
Published GitHub › -
Forecasting Fight Outcomes with Hybrid ML Models
Predicts MMA fight results from fighting style, physical attributes, and experience. Pairs logistic regression with a dropout-regularized feed-forward neural network to balance interpretability against predictive power (Stanford CS221).
Founder research (personal/academic) · project co-authored with Austin Salcedo · date not independently verified (repository is not under the Truth Computing founder's GitHub account)
Research GitHub › -
The Complete Pre-Truth Computing Research Record
The full archive of undergraduate projects and academic papers spanning AI, optimization, applied mathematics, and systems. The complete body of work that laid the foundation for Truth Computing.
Founder archive (personal/academic) · Matthew Torre · repo created 2025-12-11
Archive GitHub ›
Project and course references describe academic and personal work by Truth Computing’s founders. References to Stanford University (including course numbers such as CS238) describe coursework and research and do not imply that Truth Computing is affiliated with, sponsored by, or endorsed by Stanford University. Product, project, and technology names are used for identification only and remain the trademarks of their respective owners; their use does not imply affiliation or endorsement. Linked repositories are hosted on individual founders’ accounts and may be governed by their own license terms.
Beyond the Research
Other problems we’ve worked on.
The work above is a sample. Each founder keeps a fuller record of the problems they’ve chased, from quantum optimization and product MVPs to documentary film. If you want to see more, start here.
Matthew Torre
Engineering & product
Quantum approximate optimization (QAOA for the Traveling Salesman Problem), transformer fine tuning and applied ML, sports analytics models, and product MVPs like Demystifyd and EZRecruit, built from discovery through strategy, design, and financial modeling.
See Matthew’s portfolio ›Mark Torre
Narrative & media
Documentary film, journalism, and a custom stop motion technique built from hundreds of thousands of curated photos. Field documentaries across California capturing how communities really live. This is the storytelling muscle behind the brand.
See Mark’s portfolio ›