Industry systems and shipped builds

Exhibit A

An AI code reviewer that may only report a bug when it can hand you a test that fails on the broken code and passes on the fix, and stays silent when it cannot.

• One rule, enforced by construction rather than by a confidence threshold: no runnable proof, no comment. A deterministic, model-free flip check is the sole judge of what counts as evidence.
• Two modes on one evidence engine. Detective reproduces a bug from a stack trace; Prosecutor reviews a pull request and comments only when a flip is proven.
• Every proof is an execution-validated fail-to-pass test tied to a commit, so each one doubles as a contamination-free benchmark instance for AI4SE research.

• Python
• pytest
• Git
• LLM tool use

Tollgate

An industrial inspection agent that sends a frame to the cloud based on what a mistake would cost, not on how confident the model happens to feel.

• One inequality yields three properties at once. Routing on cost rather than confidence gives privacy (only a cropped region ever leaves the device), offline tolerance (queue and reconcile), and cloud economy from the same decision.
• Measured on real MVTec data across six categories: 0.988 accuracy at 57% of cloud-only spend, against 0.992 at full spend and 0.951 running purely local.
• Six-category robustness 0.969 ± 0.015, and a backbone ablation moves the hybrid result by only -0.024 to +0.023 — the router absorbs local-model variance. 168 tests green.

• Python
• ONNX
• DINOv2
• qwen3-vl-plus
• MCP
• Docker
• Alibaba Cloud
• SQLite

Warren

Turns a Wikipedia rabbit hole into a shareable map, where the path you actually clicked stays bright and every other link fades into context.

• Spine-and-focus graph principle: the clicked path is a thick animated edge, neighbours dim, distant nodes fade. Sidesteps the Obsidian/Roam-style hairball at 200+ nodes.
• Reads inside the map: clicking a node opens a floating burrow card anchored to it, so the map never leaves the screen.
• Reverse-engineered from the share card inward — the artifact is designed first (the Spotify Wrapped lesson), then the experience fills it.

• Next.js 16
• TypeScript
• Tailwind v4
• Supabase
• Claude Haiku 4.5
• react-force-graph-2d
• @vercel/og
• Wikimedia REST

Wynlabs

Founding engineer on an industrial copilot, running multi-agent workflows over live plant-floor SCADA, PLC, and MQTT data.

• Multi-agent workflows on per-deployment knowledge graphs (2k to 5k nodes), orchestrated through LangChain.
• Shipped 15+ client POCs end-to-end, embedded with manufacturing teams.
• Built FastAPI services and a provisioning CLI for Dockerized pipelines on AWS and Kubernetes.

• Python
• FastAPI
• LangChain
• Docker
• Kubernetes
• AWS
• SCADA/PLC/MQTT
• TypeScript

Source code is private (commercial product).

Knowledge Retrieval System for Technical Documents

A document assistant for a materials manufacturer that refuses to answer at all when it cannot cite the source it drew from.

• Citation enforcement blocks ungrounded outputs before the user sees them. That policy is what makes the system safe on regulated content.
• Hybrid retrieval: FAISS + BM25 fused via Reciprocal Rank Fusion, then a cross-encoder reranker.
• FastAPI + SSE backend, Chainlit chat UI, evaluation harness for Recall@K, nDCG@K, MRR, and latency.

• Python
• FAISS
• BM25
• cross-encoder reranking
• FastAPI
• Chainlit
• SSE
• Docker

Repository is private (academic capstone).

TruthCast

A fact-checker where AI agents argue both sides of a claim, and the verdict is written to a public ledger so it cannot be quietly edited later.

• Decomposes claims via HiSS, retrieves with Gemini + google_search, weights by MBFC source credibility (~4,000 domains).
• Pro/con debate fires only when inter-agent agreement falls below 80%. Emits one of 7 verdicts (TRUE through UNVERIFIABLE), not a boolean.
• Writes verdicts to a Solana memo for permanent provenance. Voice summary via ElevenLabs TTS.

• Next.js 14
• TypeScript
• Gemini 2.0 Flash
• Solana (devnet)
• ElevenLabs TTS
• Turso/SQLite
• MBFC dataset

Shift

One personalised climate action a day, with the energy and carbon cost of every AI call it makes shown openly rather than hidden.

• Daily actions tailored to commute, diet, live grid carbon intensity, and weather, drawn from EPA + DEFRA emissions data and 190 curated actions.
• Chrome extension and Eco-LLM dashboard track energy (Wh), carbon (gCO2), and water (mL) per Gemini prompt. Semantic caching serves repeat queries at zero added inference cost.
• Typical carbon ROI 10,000:1. Projected 12,000 tonnes of CO2 removed per year at 100k DAU.

• Next.js 14 (PWA)
• TypeScript
• Tailwind
• Groq (Llama 3.3-70B)
• Gemini
• Supabase
• Upstash Vector
• EcoLogits