Available for selected roles Instructor, Robocode UK Liverpool, UK MSc Data Science & AI
I design and ship production LLM applications, retrieval pipelines and agent products, end to end — from evaluation and backend architecture to interfaces people can trust.
The newest systems in the portfolio: agent markets, model workforces, local creator infrastructure, settlement safety and voice-first learning. Every link below goes to something you can inspect.
Newest flagship · Agent commerce
The market where agents hire agents.
Give Yuzu a goal and a budget. Agents discover sellers, bid, prove capability, negotiate, contract, execute, verify and settle, with signed receipts recording exactly what happened and which permissions were used.
A Windows creator studio that turns footage and ideas into editable video projects. Timeline edits, captions, narration, local media and real exports stay on the machine, with optional MCP and model workflows layered on top.
Exactly-once settlement for Request Network obligations through KeeperHub. Retries are allowed; duplicate money movement is not. The proof includes independent processes racing the same obligation and live Sepolia settlement evidence.
Study out loud. VIVA transcribes what a learner says, identifies the gap, grounds the correction in source material and brings the weak concept back for active recall. Built for AssemblyAI Voice Hackathon Week.
Five systems that show how I think: hired model workforces, temporal agent memory, grounded answers, memory integrity for agents, and knowledge interfaces that make complex data usable.
The intelligence resource manager under the model you already pay for. It turns a repository and a goal into a task graph, auctions each task to the cheapest model that can pass its tests, and hands a checkpoint to a replacement when a worker fails.
Temporal, provenance-first memory for AI agents on HydraDB. It keeps every claim tied to its source sentence, preserves corrections instead of overwriting them, surfaces disagreements and abstains when the evidence is missing, over web, CLI, API and MCP.
An evidence-backed workspace for asking one question across multiple work sources. Every supported claim opens back to the exact record behind it, so the answer is inspectable instead of merely confident.
A memory-integrity layer that checks retrieved context before an agent acts. It surfaces suspicious provenance, blocks unsafe execution and produces a signed integrity certificate for review.
A personal knowledge system that turns uploaded material into grounded answers and a navigable 3D relationship map. It makes the source landscape visible instead of hiding retrieval behind a chat box.
Rate limits, bad data, a worker that dies mid-task. The systems I ship are designed for the moment the traffic turns, not the demo.
I am an AI systems engineer and MSc Advanced Data Science & AI student at the University of Liverpool.
I work across the full product path: retrieval and evaluation, model orchestration, APIs, data, deployment and the interface that exposes the evidence. The goal is not an impressive prompt. It is a system that behaves clearly when the input is messy, a dependency fails or a user asks “why?”
I am currently open to UK part-time AI/software roles and internships, plus graduate opportunities from 2027.
Trace outputs to sources, evaluations or observable system state.
Make fallbacks, limits and uncertainty visible before they become incidents.
Architecture and interface are one product, not two separate demos.
Two separate authorised engagements against HydraDB. One placed top 3 in a competitive hack; the second was a structured owned-tenant assessment. Reported through coordinated disclosure; sensitive specifics withheld.
Round 01 · Hackathon Top 3 of 20+
Found an authentication / credential-handling weakness reachable without proper authorisation. Placed in the top 3 of a field of 20-plus.
Technical specifics, payloads, and affected endpoints withheld under responsible disclosure.
Round 02 · Bug bash 6 defects
Authorised read-only assessment of an owned tenant. Filed 6 reproducible defects across reliability, input validation, and API-contract correctness. Tenant isolation held throughout, with no data exposure; the core security proved sound.
Defect detail shared privately with the maintainers.
Won with QueueProof, an evidence-backed cross-source retrieval workspace where every claim opens back to its source record.
See the build →Won the Build Blitz with HydraSentry, a graph-native context-integrity harness on HydraDB.
See the build →Reached the finals with Qyntra, a self-organising wiki, built solo in roughly a day.
See the build →Entered DelOS, a delegation operating layer for coordinating multiple agents under load.
See the build →The corrected version 2 presents the architecture and analytical latency model without claiming prototype or gameplay-corpus results. It also repairs the citation error documented after v1.
Corrected v2 · Original v1 DOI archived
University of Liverpool, UK
Christ University, Bengaluru
VYN3SU1H9I3Z · Nov 2025
H4BG7UI694TN · Nov 2025
2W2FIN6EVBSL · Nov 2025
OO1BFMY72ZDB · Nov 2025
ZDO3B89P0WKZ · Nov 2025
Open to UK part-time AI/software roles and internships, plus graduate opportunities from 2027. Fastest reply is email or LinkedIn.