Public view. Some sections unlock with a personal link.
NothingItIs
AI Systems & Automation Engineer · Biomedical Sciences ahead
18, UK — and this page is all of me, not just the code. I'm headed into Biomedical Sciences, with an NHS Emergency Department volunteering placement secured and a five-year picture I mean literally: cure cancer, build the AI-powered hospital. I'm also a self-taught systems builder (since 2023, nobody in the room): I architect, AI types, both of us verify — a real production AI platform runs on my own server to prove it. Jobs, uni, medicine, projects, the life behind them — one story, and it's all here.
Simulation training on an advanced CPR doll — defibrillator, bag-valve ventilation, compressions done properly.
Highlight: a professionals' panel and hands-on diagnostic kit, back in Year 10.
Grade 3 → 8: the chemistry comeback
My lowest grade ever, right before end-of-year exams — answered with a system, not a mood.
Highlight: chemistry went from my weakest result to my strongest science.
The maths technique off the spec
Found a graph-transformation method for the area between two curves that wasn't in our exam board's specification.
Highlight: escalated by my teacher to the Head of Maths; I kept researching it on my own.
Pearson Scholarship essay
The cancer-origin story and the hospital experience in polished written form — submitted November 2025.
Highlight: the same five-year picture as this site, documented a year earlier. Consistent, not retrofitted.
problem → build → hard decision → measured result
Flagship projects
architecture, decisions, and the honest failure stories
1 · Multi-agent AI platform, run like production
Persistent assistants with memory, tools, and identity — isolated per person, rebuilt greenfield after a year, cut over with zero data loss.
Highlight: a local embedding model beat the commercial cloud one in my own regression-gated benchmark — and shipped.
2 · Real-time operations dashboard
Mission control for the platform: live sessions, costs, health, devices — event-driven over WebSockets, framework-free frontend.
Highlight: baseline CPU from 127% to under 1% — by architecture, not hardware.
3 · Android assistant client, from scratch
A phone turned full remote client — wake word, biometrics, recovery chains — on an OEM that kills background apps for sport.
Highlight: a VPN-loss interlock that disables the phone's own debug channel rather than leave it exposed.
4 · Security engineering, incidents included
OS-level credential boundaries the model process physically cannot cross, default-closed networking, audited emergency access.
Highlight: my own audit found a real cross-instance exposure path — closed same day, told here on purpose.
evidence, not vibes
Claude usage
the exact numbers my own ops dashboard computes — message-deduplicated, priced per model version, zero API calls
£15,000+ of tracked Anthropic usage, concentrated in the 2026 platform era — tracked with telemetry I built myself. Cost observability, model benchmarking, caching, and dependency-deletion are engineering habits here, not slideware. (Chart: input+output tokens by model family — the same hourly series and the same smoothed-stack renderer as my dashboard, auto-bucketed across the window; legend cost and totals are all-time, deduplicated, priced per model version. Redacted per policy — no keys, no account identifiers, no private project names.)
the receipts
GitHub, in chronological order
public repositories — every one verified clean across its full commit history
AI: Claude Code & Anthropic API (primary since Feb 2026) · MCP (heavy custom server author) · multi-agent orchestration · local embeddings · wake-word / speaker-verification models
Automation stance: code-first by choice after a hands-on n8n evaluation — built the first memory system in it, then replaced it for control over state/testing/perf/security. n8n is fully retired from my stack (July 2026) — I'd deploy it again where a client's stack calls for it.
Practices: research-before-code · delegated-agent verification · versioning with rollback everywhere · changelog-enforced releases · human sign-off on every deploy/version/deletion · incident-driven rule-making · a guides library written for AI consumption — documentation as load-bearing infrastructure
yes, actually everything
The Everything Index
every service, bot, app, mod, guide, experiment — and every utility script down to the 30-minute cron keeping this very domain pointed home. 🟢 live · 🟠 parked · ⚫ retired · 📦 archived
The AI platform core 9 entries
🟢 Platform v2 — the greenfield rewrite of the whole personal-AI platform, cut over to production with zero data loss.
🟢 Assistant #1 (primary) — persistent Claude session on Telegram: 100+ tools, graph memory, biometrics, server control.
🟢 Assistant #2 — a second, fully isolated instance built for someone close — its own OS-level world by construction.
🟢 Email tagger — a third persistent AI daemon that triages Gmail and learns from one-tap corrections.
🟢 Memory system v2 — graph memory with dedup, contradiction detection, decay — and a benchmark upset.
🟠 Platform v1 — the first generation (Feb–Jul 2026): persistent-daemon architecture, 24 tool modules. Parked as history.
⚫ The n8n workflows era — the assistant's birth certificate: echo bot → agent with memory and SSH hands.
⚫ Third instance for a friend — provisioned May 2026, decommissioned Jul 2026. Proved the platform mints instances.
📦 Instance factory — an agent that creates new assistant instances end-to-end, validated on a disposable test instance.
The v2 service constellation 7 entries
🟢 Three production instance daemons — each a hardened systemd service under its own Linux user.
🟢 Fleet watchdog — health monitor + self-heal for every daemon in the fleet.
