Production AIthat ships.

Prime Radiant

Epidemic forecasting · vintage-honest ML · best work to date

Live · dashboard + PyPI · CDC hub registration open

A CDC FluSight forecaster that predicts weekly flu hospital admissions for every US state and territory, 23 calibrated quantiles at a time, built so the numbers can be trusted: the model never sees data dated after the forecast moment — training data is checked out of the hub's git history as it existed on that date — and the code path that submits live is the same one backtested across 55 historical weeks. LightGBM ensembled with a replica of the CDC baseline; in the 2025-26 season backtest it is ahead of UMass-flusion on natural-scale relative WIS (0.609 vs 0.625) and behind it on the log scale, and it prints the two earlier seasons it loses just as plainly. 100% test coverage with the CI gate set there, every workflow gate mutant-tested, and the off-season runs itself: a shadow forecaster that arms automatically the week CDC's data feed wakes, and a watcher that files the go-live checklist the day the new season's config lands.

PREVERA Guardian+AI

Floor 2D LIDAR + cameras · on-device RF-DETR

Public research prototype · Patent pending

A privacy-first fall-detection research prototype for senior-care rooms, on a Jetson Orin Nano. A floor-level 2D LIDAR finds a person lying down from geometry alone, except end-on. For that blind spot, RF-DETR is served on the device by Roboflow Inference. On recorded camera frames the stock model found the person and could not tell lying from standing, so I fine-tuned one on Roboflow; it passed its pre-declared bars on public data, then on recorded room frames from the counter camera, while the floor camera failed one pose. Detection today, prediction next; nothing predicts falls yet. One subject, one room. The failures are published beside the passes.

Governance Drift Researcher

Agent governance · evidence-verified findings · CI gate

Live · PyPI + WeaveMind Cloud

Vendor model lifecycles move; the approved agent inventory doesn't. This scans an Azure AI Foundry catalog against a human-owned baseline and reports the drift under rules that make the report worth trusting: every finding carries its source URI, JSON field path, and content hash, and any finding whose citation no longer resolves is dropped and counted rather than published. Coverage gaps are stated, not implied. It exits nonzero on surviving drift so it works as a CI gate, and emits SARIF for Defender and code scanning. Also ported to Weft and run end-to-end on WeaveMind Cloud, where the human approval gate is structural — there is no API by which the agent can approve its own report.

Triton Kernel Lab

GPU kernels + compiler IR · study artifacts

Public · CI green on every kernel

Triton kernels — elementwise, softmax, matmul — written to be read as much as run. Each one commits its own lowering artifacts: TTIR, TTGIR, LLVM IR, and PTX, so the compiler's choices are inspectable rather than asserted. Benchmarks put the matmul against a hand-written CUDA C++ baseline and cuBLAS, with the correctness gate bounded by cuBLAS's own error against fp64 rather than an arbitrary tolerance, and softmax against torch.compile's steady state with a committed Dynamo explain artifact. The Jetson bring-up log records six failures in the order they happened, including the one an external cold read caught in the repo's own message.

GUSS

Edge inference appliance · headless + autonomous · solo build

Running · public case study

A governed autonomous agent on 7 watts. A hand-assembled, permanently headless Jetson Orin Nano serving a local Qwen3-8B at 12 tok/s that monitors its own health, restarts its own services, paper-trades a benchmarked portfolio inside code-enforced guardrails, and publishes its own dashboard — twice daily, unattended. The build's core lesson, earned across four documented model failures: code decides, the model narrates. Architecture, ADR-style decision log, and an honest build log in the public case study.

Compass-BlackBox IQ

Microsoft Foundry + MCP · solo build

Agents League · AI Skills Fest 2026

A git-backed, Markdown-native blackbox flight recorder and skill auditor for autonomous agents. It governs the layer permissions and orchestration don't — memory and competence. Every action is an immutable decision record; audit heuristics turn uncited decisions into draft skill proposals a human approves. Agents propose; humans promote. Grounds on Microsoft Foundry IQ, kept orthogonal to governed memory.

Provenance

Citation-grounded multimodal RAG · no GPU

Live

Ask a textbook a question; get an answer where every claim is checked against the exact page image that proves it. Retrieval is Cohere Embed v4 over 1,347 page images; answering and verification are Claude vision calls in a LangGraph verify→repair loop. Faithfulness 0.985, recall@5 0.864 — reported as measured, not tuned. Fully API-based and serverless.

Clinical AI Agent

SMART on FHIR + transparent RAG

Live demo

Patient-grounded, citation-traceable clinical decision support on open standards. A 5-agent pipeline fuses a patient's live SMART on FHIR record with cited evidence, emits dual citations (patient data + literature), and scores its own faithfulness with a built-in Clinical Work IQ harness. Built to fill the unoccupied square in a scan of nine frontier healthcare-AI companies: open, transparent, patient-grounded, self-proving.

FetchMerck AI

Published RAG · clinical decision support

DOI 10.57967/hf/8101

An end-to-end RAG pipeline published on Hugging Face with a DOI: ingestion, chunking, embeddings, semantic retrieval, re-ranking, prompt engineering, deployment. LLM-as-a-Judge scoring rubrics across multiple clinical query classes — the evaluation infrastructure is the asset, not the model itself.