Every number below traces to a committed benchmark, test run, or published result, losing axes included. Seventeen of these run in your browser: not a recording of a benchmark, the repository’s own code compiled to WebAssembly or executed on the committed data, so you can move the input and watch the result move with it.
strata · LSM-Tree Storage Engine
A leveled LSM-tree key-value store from scratch: checksummed WAL, MVCC snapshots, Bloom-filtered SSTables, group commit. Durability verified by a harness that tears writes and SIGKILLs the engine mid-flight; the same matrix gates every push in CI.
11,149 SIGKILLs · 2.22M acked writes, 0 lost · 50× RocksDB on sync commits
C++20libFuzzerTSanYCSB
taut · Reliable-UDP Transport + SWIM
Transport library for small messages on lossy links: SACK-based ARQ, adaptive RTO, SWIM failure detection on a lock-free epoll loop. Trades bulk throughput for tail latency, and publishes the losing axis (kernel TCP wins clean-link throughput ~27×).
p99 ~6× below kernel TCP at 5% loss · ~12× at 20% loss
C++20epollnetemlibFuzzer
tautq · Distributed Webhook Delivery
Built on taut: a coordinator-less delivery service replicating a write-ahead log by majority commit under epoch-fenced leases, with no ZooKeeper and no etcd. Building it surfaced two SWIM protocol bugs, fixed upstream. A chaos suite gates every PR.
800 jobs/s · p99 107ms · 100% completion at 0–20% UDP loss
C++20WAL replicationchaos testing
Tollgate · Multi-Tenant LLM API Gateway
Lets a team share one LLM provider key safely: personal revocable keys with rate budgets, the real credential injected server-side. Atomic Redis rate limits hold a global ceiling across replicas where per-replica limiters admit 3× the limit.
2,000 req/s, 0 errors, p99 46ms · 300 req/s global limit held across 3 replicas
GoRedisKubernetesTerraformOTel
tickvault · Full-Depth Order Book Archive
There is no free source of full-depth crypto order book history you can trust: everything free is top-of-book, resampled, or carries undocumented gaps. This records six venues to Parquet and publishes its own coverage, so a window you can't trust says so instead of interpolating.
6 venues · reconstruction to any instant · every gap published, not filled
RustWebAssemblyParquetArrow
06
Verify-then-Escalate Model Cascade
Two-stage routing (Opus 4.8 primary, GLM 5.2 FP8 rescue on verifier failure) over 100 frozen EvalPlus tasks, ranked by $/solved-task. Beat every single-model baseline; the gain traced to complementary failure modes, not a uniformly better model.
94/100 vs best single model's 90 · 47.1% cheaper · published by GMI Cloud
PythonEvalPluscost accounting
published result ↗
07
Mint · Lead Intelligence Platform
End-to-end lead intelligence at GMI Cloud: identity resolution across Apollo and Exa evidence chains, LLM ICP tiering with human-review gates and audit trails, and a 16-tool conversational agent over the lead database.
18K LOC · 45 endpoints · 251 tests · 51 qualified leads from a 367-person triage set
FastAPIPostgreSQLReactLLM classification
internal @ GMI Cloud · demo coming soon
08
Winnow · LLM Prompt Compression
Voice-first compression on two orthogonal axes: token-space (LLMLingua + self-implemented AttentionRAG behind an order-preserving merge) and model-space (~4-bit TurboQuant KV cache + LCLM latent compression), so the savings multiply.
~60× KV cache reduction · F1 0.437 vs 0.376 fp16 baseline on LongBench
PyTorchModalDeepgram
repo ↗
09
Pylon · Counter-UAS Sensor Mesh
Built at the Bow Capital Defense Hackathon (UCSD). Anomaly-based RF detection over dual SDR backends, a UDP gossip mesh, and a 3D operator dashboard in Next.js + deck.gl.
Catches novel emitters a signature-library baseline misses 100% of · mesh converges ~2s
Pythonscikit-learnSDRdeck.gl
repo ↗
10
Vigil · Autonomous AI On-Call Engineer
Built at the Loop Engineering Hackathon: detects, diagnoses, and remediates production incidents end-to-end while holding zero standing credentials: single-use, scope-bound grants behind a Pomerium identity-aware proxy, with policy-denied escalation and a full audit trail.
5 services · zero standing creds · single-use 60s-TTL grants
TypeScriptPomeriumDockerAkash
repo ↗
11
LOCUS · Voice-Driven SAR Command Center
Search-and-rescue ops: LLM agents profile the missing person, a Monte Carlo engine turns hypotheses into a live probability heatmap and team assignments, and the coordinator drives it all hands-free by voice.
150 walks/hypothesis × 4 ISRID profiles · 1.5s sim tick with zero LLM calls in-loop
TypeScriptConvexMapbox GLWeb Speech
repo ↗
Studies you can run.
Each of these started as a claim I could not find a number for, so I built the smallest
thing that would settle it. The page runs the repository's own code in your browser, on the committed
data, so you can move the input and watch the finding hold or break. Eleven of them, all reproducible.
