OptimaTech Labs

OptimaTech Labs · engineering journal

Inside the Lab

Real systems. Real experiments. What happened when we built them.

Less whitepaper, more wrench.

We’re a small engineering lab in Rhode Island. We work on artificial intelligence, autonomous systems, simulations, cryptographic verification, and the occasional thing that doesn’t fit neatly into any category.

We don’t always know whether an idea will become a product. That’s part of the point. Inside the Lab is where we open up the work: the architectures, the experiments, the data, the decisions, the unexpected results, and the things we had to rebuild because the first approach didn’t work.

We show the actual technology wherever we can: working demonstrations, measured results, the hardware it ran on, and the limits that matter. No imaginary roadmaps presented as accomplishments. No performance numbers without context. No pretending an experiment succeeded when it didn’t.

Built here. Tested here. Explained here.

Field note 008 · Private Model · Encrypted AI

Not Us. Not Our Servers. Not Even the Model.

We trained an AI model on encrypted data its host never saw. The host couldn’t read the data, the questions, the answers or the model itself. Only the owner’s key opens them.

0 readable words · 242 sealed training steps · post-quantum

Read the field noteTry it live
7b 05 10 bc 6f 73 03 44 68 52 a7 4a 70 6c 04 cf 47 a8 2a 15 0e 56 00 f2 99 2c 9c 5c a6 07 00 7e ed 1b 56 b0 21 d0 03 d8 4a cc 30 fb db 76 02 3f ec bc 47 63 f7 12 04 10 74 a4 1f e9 8c 4d 02 21 bb 39 68 8e c4 cd 00 99 7a 46 4b ff 8c 3b 02 3f 35 7b ad 9d 72 e0 01 39 85 45 fb e3 11 d2 03 5b 11 4b ae dc 32 85 04 27 83 bf 96 cd 34 e5 07 7b ab 03 9a 97 7f 98 00 8e 35 72 68 61 c6 d8 04 f2 7f 41 e1 eb 94 e9 03 f7 8a 56 f8 f7 6f c8 07 71 47 6e d8 3e b6 9b 00 da 8e 9c 53 fd be cf 07 c7 … What our server holds · hover or tap

Lab log

  1. Scrapped Agent Desk before release, for bigger and better things. WORLDSTATE takes its place. WORLDSTATE
  2. Agent Desk: the dashboard we run 200+ agent-built products from. Agents propose, a human decides. Field note 010
  3. Fixed the long-run fault: noise was piling up in the unused half of each encrypted number. Training now runs clean past the step that used to break. The fix
  4. Trained an AI model on encrypted data its host never saw: 242 steps on one GPU. Field note 008
  5. Moved encrypted training from CPU to GPU: about 12x faster per step. The numbers
  6. Caught a fault on our GPU encryption path during a long run and traced it to one operation. The fix comes before bigger runs. What it was
  7. Countersign: a pre-funding invoice check for lenders, in development. Countersign
  8. Kicker’s decision model went to production. Optima Prime signs every batch of hands. Field note 001
  9. Kicker Live: six characters play cards at watching speed. Watch it
  10. THREADPULL: an AI investigator that has to show its evidence. Try the demo

The field notes

  1. 010 · Agent DeskManagement of ScaleScrappedMore than 200 products built by agents since May, and the one screen where agents propose and a human decides.
  2. 009 · The Conflict LedgerA Forecaster That Has to Show Its WorkPublishedCheckable questions, linked sources, officials’ claims judged against the record, and a public score against prediction markets.
  3. 008 · Private ModelNot Us. Not Our Servers. Not Even the Model.PublishedTraining an AI on encrypted data its host never sees: post-quantum encryption, a GPU, and the controls.
  4. 001 · KickerWe Taught a Small AI Model to Play a Character, Using 120 Million Simulated HandsPublishedSimulation, synthetic data, fine-tuning, WebAssembly, replay and signed records.
  5. 002 · Optima PrimeWhat If Every AI Answer Came With a Receipt?ComingSigned inference, model fingerprints, audit trails and replay.
  6. 003 · THREADPULLWe Built an AI Investigator That Has to Show Its EvidenceComingSaved sources, exact quotations, contradictions and gaps.
  7. 004 · Synthros and NonplayerWhat Happens When AI Characters Have Lives Outside the Conversation?ComingMemories, motivations, autonomous behaviour and simulated populations.
  8. 005 · Optima PrimeWe Taught a Small AI Model to Know When It Doesn’t KnowComingHuman corrections, retraining and off-topic refusals.
  9. 006 · TerritrixFinding Gambling Operators That Don’t Want to Be FoundComingDomain discovery, game identification and evidence for regulators.
  10. 007 · feel-smart-llmWe Built a Transformer in NumPy to Understand Every Moving PartComingModel internals, hand-built backpropagation and a browser demonstration.

Scrapped

Projects we stopped. Their pages stay up as the record.

  1. 23 Oct 2026Agent DeskScrappedA dashboard for teams of AI agents: they propose, you approve. Never released. We moved on to bigger and better things: WORLDSTATE. Field note 010