The full list
All projects.
Everything, in one place: systems, ML, agents, studies, and hackathon builds. Every number traces to a committed benchmark, test run, or published result, and seventeen of these run in your browser on the repository’s own code. Builds marked team were made with other people.
- strataLSM-tree storage engine in C++20: WAL, MVCC, group commit, crash-verified durability11,149 SIGKILLs · 2.22M acked writes, 0 lost · 50× RocksDB on sync commitsrun it ↗
- tautReliable-UDP transport + SWIM membership for lossy links, losing axis publishedp99 ~6× below kernel TCP at 5% loss, ~12× at 20%run it ↗
- tautqCoordinator-less distributed webhook delivery on taut: WAL replication, epoch-fenced leases800 jobs/s · p99 107ms · 100% completion at 0-20% lossrun it ↗
- TollgateMulti-tenant API gateway for shared LLM keys: atomic Redis limits, OTel, Kubernetes2,000 req/s, 0 errors, p99 46ms · global limit held across 3 replicasrun it ↗
- Model CascadeVerify-then-escalate routing, GLM 5.2 then Opus 4.8, over 100 frozen EvalPlus tasks, published by GMI Cloud94/100 vs 90 for Opus 4.8 · 47.1% cheaper than Fable 5post ↗
- MintLead-intelligence platform at GMI Cloud: identity resolution, LLM ICP tiering, 16-tool agentlead-research time cut 90% · enrichment spend cut 75%internal
- WinnowQuantized KV cache with a fused Triton kernel, on top of a voice-first prompt-compression hackathon build34× faster dequantization for 0.47 points of LongBench-Erepo ↗
- PylonteamCounter-UAS RF sensor mesh: anomaly detection, dual SDR backends, 3D operator dashboardflags novel emitters a signature library cannot know about · mesh converges ~2srepo ↗
- VigilteamAutonomous AI on-call engineer: zero standing credentials, single-use scope-bound grants5 services · 60s-TTL grants · full audit trailrepo ↗
- LOCUSteamVoice-driven search-and-rescue command center: Monte Carlo heatmap, LLM tip reweighting150 walks/hypothesis × 4 ISRID profiles · 1.5s tick, zero LLM in-looprepo ↗
- AgentBenchEval framework for LangGraph agents: cost-adjusted accuracy, Ray Tune HPO, CLI + leaderboard0.964 acc · n=55 · $0.042 total spendrun it ↗
- Checkit Health7-stage misinformation-detection pipeline over 4 live social APIs, cost-tiered LLM path15 req/min free tier heldrepo ↗
- ArgusAgentic training monitor: SPC detection across 4 failure modes, bounded self-correction49% of anomalies auto-resolvedrepo ↗
- SitRepteamTeam project-management app: led 11 engineers through a 10-week Agile SDLC on Cloudflare11 engineers · 5 sprints · shipped livelive ↗
- ChainSenseteamEthereum behavioral analytics: 16-feature wallet matrix, 9 archetypes, anomaly flagging403K txns · 59K wallets · 591 anomaliesrepo ↗
- RepoMapVS Code extension mapping any repo as an interactive dependency graph with AI explanations6 languages parsedrepo ↗
- AgriShieldteamWildfire risk + firebreak optimizer: ELMFIRE ensembles score 24 candidate layoutsrepo ↗
- UrbanSound8KCNN-free audio classification: 487-dim engineered features, wavelet scattering, leak-free CV72.9% top-1 · 0.51ms/clip inferencerepo ↗
- Helmet IdentificationResNet-50 pipeline detecting riders and classifying helmet use; published via Lumiere94% acc · 0.92 mAP · <50ms inference · top 10% of ~200 paperspaper ↗
- tickvaultFull-depth crypto order book archive in Rust: six venues to Parquet, reconstruction to any instant, coverage published rather than interpolatedevery gap published, not filledrun it ↗
- amacControl plane for the AI coding agents on my Mac: every session and automation on one page, a work queue with fencing200 SIGKILLs · 8,000 acked events · 0 lostrepo ↗
- decode-anatomyWhere an LLM decode step actually goes on an RTX A6000: kernel time, gaps, and launch overhead separated38.3% of a batch-1 step is not kernel work, and batching removes itrun it ↗
- ddp-crossoverWhere a second GPU stops paying on 2x T4 over PCIe, measured across model sizesspeedup crosses one going up, the opposite of the usual advicerun it ↗
- noisy-rag-benchWhat OCR noise does to retrieval versus to answers, on real scanned pagesat 5% error recall is unchanged and half the answers are already gonerun it ↗
- open-refusal-steeringOpen reproduction of feature-level activation steering, with a coherence check attachedat 100% response rate every behavioural check passes and the output is token souprun it ↗
- memharnessWhat a read-once-answer-forever memory system actually costs, priced per turnthe one read cost 12.7x all twenty questions combinedrun it ↗
- actiongateA confidence gate and tamper-evident audit trail between a meeting transcript and a CRM writegating at 0.65 turns 18 wrong writes into 2, keeping 19 of 22 genuine onesrun it ↗
- icp-benchmark120 real YC companies hand-labeled on three axes, with a scoring harness for lead qualificationenrich-everything is 25% precision; a rules filter is 93.8% and misses halfrun it ↗
- promptgateWhat happens when a published production prompt states its contract four times and contradicts itself oncethe parser reads a reject as silence and the gate treats silence as approvalrun it ↗
- pms-raceSixteen computer-use agents racing for one hotel room, with and without a transactionthe legacy path books it 93 times; the obvious fix does nothing for lost updatesrun it ↗
- flightriskDeposit-flight risk across every US bank, 2015 to 2026the risk is U-shaped: a large inflow predicts a drawdown about as well as an outflowrun it ↗
- apparatus-compsA comparable-sales estimator for fire apparatus, back-tested against real saleshalf the estimates are off by more than 36%run it ↗
- Project LiftoffIncubator management platform: 6-tier RBAC via Supabase row-level security, 13 tables100+ users onboardedprivate