limit-order-book — one thread owns all book state, orders arrive through a lock-free MPSC ring, and the match path is built so the garbage collector has nothing to do.

The baseline is a fair idiomatic Java implementation held to the identical contract by a shared test suite, not a strawman. The trick is intrusive structures: resting orders are their own FIFO queue nodes, price levels are their own AVL tree nodes, so there's no wrapper object to allocate.

warpline — fused RMSNorm + residual at 64×8192:

The 14x is the fun number. The 1.52x is the honest one, because torch.compile is what anyone serious would actually compare against. It's the only repo I own where the "claimed but not yet verified" list is completely empty.

railyard — a SQLite control plane hands fenced leases to a worker pool, so a resurrected zombie worker cannot write over its own replacement. Roughly 60% of the codebase exists purely to prove the scheduler is deterministic.

It publishes a documented miss on 2 of its own 5 release targets. evidence_valid and qualified are separate fields on purpose, so a clean miss cannot be laundered into a pass.

build-sleuth — your build goes red. The expensive question isn't how do I fix this, it's is this even my fault: a real regression, a flaky test, and a runner outage look identical in the log.

It reads the failing run, classifies it, ranks the culprit file, drafts a patch, then applies that patch and reruns the original job inside a container. It only opens a draft PR if the build actually goes green.

The eval harness is itself tested: every case runs against the maintainer's real fix and a deliberately corrupted copy. That negative control caught a bug where a fuzzy patch fallback was accepting patches git would reject.

The first real failure it ever saw was a live GitHub Actions outage, and it declined to write a patch. That was the correct answer, and the behaviour I designed for.

In progress: exposing the pipeline as an MCP server over stdio.

medi-agent (team project — I own the backend agent, model routing, RAG, and auth layers)

pgvector over FDA DailyMed drug labels, 768-dimension embeddings on an ivfflat cosine index. The design decision I like is the SQL metadata prefilter that runs before ranking, restricting candidates to the patient's own active medications, so the vector search physically cannot surface a label for a drug they aren't taking. Below a 0.72 similarity floor it returns nothing and the prompt branches to make the model admit it lacks grounding.

Triage safety is the best-tested part: 31 golden routing cases in English and Spanish, asserted with the LLM forced to fail, so the deterministic fallback carries the emergency path alone.

penstock — streaming passthrough, provider fallback chains, half-open circuit breaking, per-tenant spend caps. One Go binary, 1.02 ms of added latency per request against LiteLLM's 12.14 ms, measured over 2,400 samples per arm with raw output committed.

It ships a semantic cache that is disabled by default, because a 257-probe labeled study found no safe similarity threshold: opposite-meaning questions outscored genuine paraphrases (AUC 0.39), and 43.9% of hits at the shipped floor answered the inverted question. Killing my own feature with my own evidence is probably the most useful thing in that repo.

millrace — PostgreSQL → Debezium → Kafka → Spark → Parquet → dbt → DuckDB. The publication gate at the end recomputes row counts, per-column SHA-256 checksums, partition counts, and business aggregates against an independent source-of-truth path. 39 of 39 checks have to agree or the published views don't swap.

flow-engine — a TypeScript scheduler resolves a JSON dependency graph and farms steps to Python workers over Redis. Three-colour DFS cycle detection distinguishes a diamond (a node legitimately reached twice) from a real back edge. Every execution embeds a frozen snapshot of the workflow it started with, so a long-running job can't change shape mid-flight because someone edited the definition.

cross-foot — routes dealer statements (PDF, CSV, XLSX) to deterministic extractors or a vision model depending on file signature, then scores how much it trusts every individual field it read and sends only the shaky ones to a human review queue.

Per-family Platt scaling is fit over the logit rather than the raw confidence, so the identity transform stays expressible and a family that's already calibrated survives being fit unchanged.

A test walks the AST of every module and fails the build if scoring code can even reach the answer key. It exists because a weaker, import-only version of that guard once let ground truth leak into the review database.

acquit — most test-selection tools skip tests because a heuristic thinks they're probably fine. acquit attaches a machine-checkable proof to every skip, and when it can't prove something, it runs the test. Shipped to PyPI and the GitHub Marketplace.

Telephony audio comes off Twilio at 8 kHz mulaw straight into the Deepgram Voice Agent socket, with six function-calling tools firing mid-conversation.

The sharpest thing in that codebase came from a bug. The API has two steering channels: one silently updates the agent's instructions, the other makes it speak. Confusing them meant internal guidance text got read aloud to a patient mid-call.

They're now strictly separated, with forced speech queued behind a speaking-state flag that keeps only the newest message and discards the rest, because stale conversational context is worse than none. There's also a prefix filter that drops steering strings if the model echoes them back, so they never land in the saved transcript.