AI systems engineer building agents that can prove what they did.
Liverpool, UK · MSc Advanced Data Science & Artificial Intelligence · University of Liverpool
The market where agents hire agents.
Yuzu takes a goal and a budget and runs the deal end to end: discover → bid → prove → negotiate → contract → execute → verify → settle. Seller claims are treated as evidence to test, permissions are scoped before work begins, and every completed deal can return a signed receipt describing what happened.
🏆 SharedOS Hackathon · Judges’ Pick · 2026
Live market · Source · Grant map · Manifest
| Result | Project | Event | What was built |
|---|---|---|---|
| 🏆 Judges’ Pick | Yuzu | SharedOS Hackathon | Agent-to-agent market with proof, scoped authority and settlement receipts |
| 🏆 Winner | Leverage | RocketRide × SCU Buildathon | Model workforce manager with per-task auctions, test verification and cognitive handoff |
| 🏆 Winner | QueueProof | HydraDB × Connectors Hackathon | Evidence-backed cross-source retrieval with claim-level provenance |
| 🏆 Winner | HydraSentry | HydraDB Build Blitz | Memory-integrity firewall that blocks poisoned context before an agent acts |
No scoreboard inflation here: finalist and participation results stay separate from wins.
| Project | What it demonstrates | Links |
|---|---|---|
| Yuzu | Agents discovering, proving, contracting, executing and settling work with signed receipts | Live · Source |
| Leverage | Per-task model auctions, zero-budget routing, cognitive handoffs and repository-test verification | Live · Source |
| VYREALM | Local-first creator studio for editable video projects, real exports, captions, narration and optional MCP/local-model workflows | Source |
| ReqKeeper | Exactly-once settlement for Request Network obligations through KeeperHub, including concurrent-worker and live-chain evidence | Live · Source |
| VIVA | Voice-first active recall using AssemblyAI Dictation + live streaming with source-grounded feedback | Live · Source |
| Lacuna | Temporal memory, exact provenance, contradiction handling and explicit abstention across web, CLI and MCP | Live · Source |
| QueueProof | Cross-source work retrieval with claim-level citations and inspectable evidence | Live · Source |
| HydraSentry | Agent memory integrity, poisoned-context detection and signed integrity evidence | Live · Source |
| Kodro | Offline robot coding and visual kinematic simulation for pupils, teachers and beginners | Live · Source |
| Cherry | User-owned skills, memory, missions and verification for AI agents | Live · Source |
- agent markets, orchestration and tool-using workflows;
- memory, retrieval, provenance and evidence graphs;
- evaluation systems that expose failures instead of hiding them;
- MCP servers, CLIs, connectors and cross-surface contracts;
- full-stack AI products with TypeScript, Python and durable data stores;
- reproducible demos, tests and technical documentation.
Languages TypeScript · Python · JavaScript · SQL
Frontend React · Next.js · Vite · Tailwind CSS · Three.js
Backend Node.js · FastAPI · REST · serverless functions
Data PostgreSQL · MongoDB · HydraDB · Redis · vector search
AI systems RAG · agent workflows · MCP · evaluation · provenance
Delivery GitHub Actions · Vercel · AWS · GCP · Docker
I prefer proof before claims: a working route, a reproducible command, a failing test that becomes green, a signed receipt, or an evidence artifact somebody else can inspect.
I document limitations alongside capabilities. If a system cannot support an answer or safely complete an action, it should say so instead of manufacturing confidence.
I am especially interested in the infrastructure underneath useful agents: how they discover one another, decide who should do work, carry memory across tasks, constrain authority, verify outputs and move value without duplicating side effects.
That is the line connecting Yuzu, Leverage, Lacuna, ReqKeeper, QueueProof and HydraSentry.
For AI/ML engineering, applied-AI collaboration or open-source work: