Skip to content
View tritsystem's full-sized avatar

Block or report tritsystem

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please donโ€™t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this userโ€™s behavior. Learn more about reporting abuse.

Report abuse
tritsystem/README.md

Hi, I'm Gavin ๐Ÿ‘‹

Independent researcher working at the edge of unconventional computing โ€” ternary and neuromorphic architectures, physical (phononic) reservoirs, and what happens when you put those ideas inside real software and games.

I build things, measure them honestly, and write down the negative results too. The discipline behind all of it is written down as The gbranaa4-hue Method โ€” measure don't infer, pre-register the verdict, and report the "no" as carefully as the "yes".

LinkedIn

๐Ÿ”ง Open-source contributions (live)

Bugs found in other people's libraries and fixed upstream โ€” mostly a recurring dtype-safety class (a state/gate tensor built with device= but no dtype=, so float16/bfloat16 input gets silently upcast to float32), plus one new feature. This table lists only what has landed: pull requests I authored that were merged, and bugs I reported that a maintainer then fixed. It is rewritten daily straight from the GitHub API by a workflow in this repo โ€” every row is re-verified against a merged commit on each run, not hand-typed. The full audited record, including the open PRs still in review and the honest negatives, is in research-portfolio/oss.

8 merged ยท 4 reported & fixed upstream ยท refreshed 2026-09-07 12:44 UTC

Repo # What Status
celery #10571 Fix Signature.clone() sharing kwargs with the original signature Merged 2026-09
kornia #4319 fix(feature): register RenderingDeFMO's times as a buffer so .to() ... Merged 2026-09
kornia #4303 fix(geometry): bbox_to_mask3d intersects axis ranges for a full/ove... Merged 2026-09
kornia #4299 fix(geometry): finite gradients for solve_cubic at its acos boundary Merged 2026-09
spikingjelly #744 Fix MSTDPLearner initialising eligibility without a dtype Merged 2026-09
spikingjelly #750 Add RAFNode: resonate-and-fire neuron Merged 2026-09
kornia #4210 fix(augmentation): auto-augment ops preserve float16 / bfloat16 inp... Merged 2026-09
spikingjelly #743 Move tensor reset values with the module in MemoryModule._apply Merged 2026-09
Pillow #9963 Image.copy() aliases mutable .info values (e.g. list-valued "Commen... Reported; fixed upstream by Andrew Murray (#9964, merged 2026-09)
aiohttp #13634 CookieJar.update_cookies() aliases a caller-supplied Morsel instead... Reported; fixed upstream by Sam Bull (#13637, merged 2026-09)
scipy #26095 BUG: scipy.optimize.milp() silently mutates the caller's options di... Reported; fixed upstream by j-bowhay (#26097, merged 2026-09)
snntorch #430 surrogate.LSO() raises TypeError โ€” wraps the wrong autograd Function Reported; fixed upstream by maintainers (#418, merged 2026-08)

๐Ÿš€ Production infrastructure

Real, deployed, live services โ€” not demos.

  • OBSERVE Search API (live product โ€” source) โ€” a hosted, pay-per-query semantic code search API for AI agents, plus a real ACP (OpenAI/Stripe) and Google UCP-compatible agent-commerce discovery and reputation layer. Read both protocol specs directly before building, since neither one actually defines cross-merchant discovery โ€” that gap is the product. Real Stripe billing, a live production deployment (Caddy, TLS, a reverse-proxied host), and a two-sided reputation system where trust is earned agreement between disconnected buyer/seller keys instead of a self-report. Same honest-negative-results discipline as the research below: three retrieval techniques borrowed from a competing tool were tested against its own corpus and measured net negative, documented instead of dropped quietly.
  • mcp-gateway (private) โ€” a self-contained, tamper-evident audit log (hash-chained JSONL) plus rate limiting for local MCP tool servers, wired into both OBSERVE's and Spikeling's MCP tool surfaces so every tool call is logged and rate-bounded, single-operator scope, no cloud dependency.
  • server-guard โ€” real-time server monitoring and alerting, benchmarked head-to-head against Datadog, Wazuh, PagerDuty, and Netdata on sourced, cited pricing rather than marketing claims.
  • spikeling-os โ€” a from-scratch x86_64 kernel whose task scheduler is driven by a spiking neural network โ€” the same Spikeling runtime below, running as real kernel-level control logic instead of a simulation of one.

๐Ÿ”ฌ Research

Ternary & neuromorphic computing

  • 012-trit-search โ€” ternary-computing research, and its shipping product OBSERVE: local, private semantic code search (desktop GUI / CLI / MCP server) with a one-line install. Nothing leaves your machine. Now includes a chunk-provenance/lineage layer (tracks a chunk's edit history across commits), hybrid (lexical + semantic) search, and an incremental indexer so re-indexing a repo only touches what changed.
  • methodlm โ€” a verifiable causal-reasoning harness: pre-registers every test, runs real backdoor adjustment (with a robustness value), and keeps an honest ledger so any LLM must prove its causal claims. A new REFUTE tool runs DoWhy-backed causal refutation (placebo treatment, random common cause, subset validation) against every accepted claim, and a real-world benchmark suite now covers 10 cited public datasets plus this portfolio's own live server-guard telemetry. Optional ternary second witness via the same tritkit readout as OBSERVE.
  • Spikeling โ€” a DSL + runtime for spiking neural networks that runs the same .spk brain on Python, C, Verilog, and Godot backends. Ships a Godot game-AI plugin, an MCP server exposing the runtime as agent tools, and a real hardware sensor-adapter layer (acoustic, system telemetry, video confirmed against physical devices).
  • llama-demo โ€” genuinely free (Ollama, $0 API cost) local-model experiments applying methodlm's causal-reasoning discipline to LLM tooling: a local coding-agent dispatcher, model vet/quantize/finetune scripts, and honest negative results (a local 7B model failed to fix a real bug even when handed the exact correct lesson).

Phononic / MEMS reservoir computing

  • quasicrystal-mems-reservoir โ€” finite-element study of quasicrystal-perforated MEMS resonators, and what their mode structure can (and cannot) compute as a physical reservoir. Includes the paper.
  • symmetry-selection-rule โ€” a symmetry selection rule for computation in those reservoirs.
  • topological-phononics โ€” does topological structure make an analog (SSH) reservoir tolerate a dead element? A pre-registered, honestly-scoped simulation study โ€” boundary conditions and negative results reported as carefully as the positives.

๐ŸŽฎ Games (built on the same spiking-brain engine)

Both are in active testing โ€” expect bugs, and some builds may not be playable end-to-end.

  • tribe โ€” a survival/RTS sim where every NPC (tribe members, animals, rival AI) is driven by a small spiking neural network instead of a behavior tree.
  • horde-defense-beta โ€” a 3D horde-defense game with a deck-building creep system.

๐Ÿ› ๏ธ Applied builds

Practical prototypes for real hardware, held to the same standard as the research above: verify it actually works before calling it done.

  • doorcam โ€” a regular USB webcam becomes a driveway monitor: YOLOv8 person/vehicle detection restricted to a hand-drawn driveway boundary, Discord webhook alerts with snapshots, and a live MJPEG feed built for screen-sharing into Discord via its native Go Live (not a bot faking video, which would violate Discord's ToS).
  • pond-health โ€” predicts pond water-quality problems (oxygen crashes, algae blooms) before they're visible and recommends organic fixes. Transparent trend-based prediction rather than a black-box model, since there's no real historical sensor data yet to train one honestly. Every run logs to SQLite, visualized in a Grafana dashboard whose color thresholds are generated straight from the same code the app alerts on. A second detector runs the real Spikeling engine (below) as a spiking neural network alongside the trend predictor โ€” honestly measured against it rather than assumed to be better: it never false-alarms, but gives no early warning, the opposite tradeoff of the trend predictor.
  • sensor-duo โ€” the trend + spiking + SQLite + Grafana pattern above, extracted into a reusable toolkit after building it by hand twice. Generic over any named numeric channel (a temperature, how long someone's lingered in a camera frame, anything with a float), with a normalized schema so a new channel never needs a migration, and a runnable end-to-end demo proving it outside the pond-specific context it came from.

๐Ÿ“– Study guide

  • the-playbook โ€” a self-study reference across this whole portfolio, taught from scratch, every claim a real measured number pulled from the corresponding repo.

The through-line

Two research threads โ€” neuromorphic and phononic โ€” that keep testing the same question: when does structure/symmetry actually help computation? That question doesn't stay theoretical: the same neuromorphic engine drives two games and a from-scratch kernel scheduler, and the same discipline โ€” measure it, verify it, report what actually happened, negative results included โ€” carries all the way through to production software people can actually pay for and depend on, not just research code that stops at a benchmark script.

Pinned Loading

  1. 012-trit-search 012-trit-search Public

    Ternary computing research + OBSERVE: a local, private semantic code-search tool (desktop GUI / CLI / MCP server) with a one-line install. Nothing leaves your machine.

    Python

  2. methodlm methodlm Public

    Verifiable causal-reasoning harness: pre-registers tests, runs real backdoor adjustment with a robustness value, keeps an honest ledger so any LLM must prove its causal claims. Sibling to OBSERVE.

    Python

  3. observe-api observe-api Public

    Hosted, pay-per-query semantic code search API for AI agents, plus an ACP/Google-UCP-compatible agent-commerce discovery and reputation layer. Live product, real Stripe billing.

    Python

  4. quasicrystal-mems-reservoir quasicrystal-mems-reservoir Public

    Finite-element study of quasicrystal-perforated MEMS resonators and what their mode structure can (and cannot) compute as a physical reservoir. Includes the paper.

    Python

  5. Spikeling Spikeling Public

    A neuromorphic DSL + runtime for spiking neural networks: write a .spk file, run it on Python, C, Verilog, or Godot backends. Includes a Godot game-AI plugin.

    C

  6. symmetry-selection-rule symmetry-selection-rule Public

    Lab report: a symmetry selection rule for computation in phononic / quasicrystal MEMS reservoirs.