LCVDB
ValidationA ternary vector database. Sharding lifted held-out recall from 92% to a clean 100%.
Bare-metal machine intelligence.
We surface the latent intelligence in the silicon you already have.
Every hard problem has structure in it. We research until we find it, then build it directly into small, deterministic systems: ternary signatures, lookup tables, a few kilobytes of C. The result runs on a microcontroller, gives the same answer every time, and holds up in front of an auditor.
Four of the fourteen, each with its stage and the number it stands on.
A ternary vector database. Sharding lifted held-out recall from 92% to a clean 100%.
Ternary lookup-table compression for language models: 3.01× smaller, faster than the original, and lossless.
A ternary soft-silicon operating system for ESP32-class chips. It is the adaptive layer that runs beside Flynn's certifiable core.
Image classification with zero learned parameters: ternary signatures, an inverted index and integer lookups.
Flynn left the lab in 2026 and now operates as FlynnComm.
An equipment-health monitor in 8,480 bytes of C99. It learns the machine it watches, locks its threshold, and runs bit-identically on any silicon, entirely on the device. FlynnComm is in discussions with industrial and aerospace buyers.
Essays argue the thesis, lab notes track the work, and reports show the results with their baselines. The first three are being written now.
Sponsored research on constrained and safety-critical compute. You get the lab on your problem, results as they land, and first look at anything that graduates.