Architecture
Artist × scientist × engineer — taste for what's good, method for what's true
MODEL CARD · HUMAN WORKER · GENERAL-PURPOSE
AQ-1 · REV 2026.07.12-2 · LONDON, UK
Aidan Quigley
AI systems engineer focused on knowledge-centric architectures
Builds AI systems that treat knowledge as the first-class citizen — represented faithfully, transformed without loss, transferred cleanly between models and people, composed into judgments better than any single source, then evaluated hard enough to prove it. Not chasing bigger models — chasing the principles by which knowledge moves between intelligences.
quigleybits@proton.me· github.com/Quigleybits· quigleybits.work· London, UK
Modalities
Python · TypeScript · Go · Postgres/pgvector · MCP · Next.js/Expo
Intended use
AI systems architecture · agent platforms · research engineering (ML × systems)
Known behaviours
Ships with decision logs · deletes what doesn't earn its place · claims only what evals prove
The through-line — one question underneath every project
How should knowledge be represented, transformed, transferred, and composed across intelligent systems — and how do we prove it worked?
REPRESENTATION
expertise in a usable form
TRANSFORMATION
one form into another, no loss
TRANSFER
models telling models
COMPOSITION
many specialists, one judgment
EVALUATION
proof that it worked
Research programme — Logia · hypotheses under test, not claims
H1 · LoRA vs prompt
Can a LoRA beat retrieval plus a strong prompt from the same corpus? Stable added value, or it doesn't earn its cost.
H2 · Conceptual adaptation
Stable conceptual relationships — or just vocabulary and style? Novel questions, held-out sources, identical evidence.
H3 · The interface
What representation carries learned expertise to a stronger model? The interface controls fidelity — it is not plumbing.
Compositional experts with explicit interfaces — not "which expert activates?" but "how is expertise composed once selected?" Evals precede training; a negative result is a valid finding.
Capabilities — strongest first
Retrieval & matching systemsCORE
FAISS + BM25 + web + graph into judged pools; embeddings proven by A/B.
Evaluation engineeringCORE
McNemar gates, frozen baselines, blind judging — evals as gates, not demos.
Agent orchestrationCORE
Planner → implementer → verifier → reviewer → critic; constitutional loops.
Protocol & tooling design
MCP servers in TypeScript; a published CLI conformance standard; npm skills.
Full-stack shipping, secure and fast
Next.js · Expo · Supabase · Firebase; RLS + bcrypt; ~160→10 ms round-trips.
Evaluations — evidence from live systems
Hymn_core — five signals, one judgmenthymncore.net
1,350 hymns over a 55,000-paragraph corpus, judged source-blind. 234 fixes live, 0 hallucinated refs; recall ~5× (p = 0.012).
2nd_brain — memory with source fidelity2ndbrain.website
Personal RAG, daily production: pgvector + Voyage, MCP server, Telegram capture — free tiers.
MCLIP — one contract across boundariesmclip.dev
A CLI conformance profile for MCP: normative spec, security model, 30+ fixtures, Go reference.
Shipped fleet — ideas taken to live
12+ live products — Assimilax (iOS App Store), scosig.com arXiv pipeline, kanban.website, cctts, claude-skills on npm.
Judgment calls — decisions and catches the evidence backed
◆
The judge selects; scores only gather.
Retrieval supplied 81% of Hymn_core's candidates but only 0.8% of final picks — so the LLM judges the whole pool source-blind, never the top score.
◆
Don't take the agent's word for it.
Eight logged catches where one human challenge beat an AI agent's confident answer — latest: an unvalidated "LoRA hosting dominates cost" claim, reversed by a price check. Narration is not evidence.
◆
Gate against Goodhart.
A regression gate cut a blind panel's "~50% improvable" to ~23% real; two blind judges must agree before a batch ships — over 20% bad-rate pauses it.
Training data — where the judgment comes from
Pre-training
BSc Biology — built a noise-tolerant plant-ID engine with error-aware trait matching; retrieval and eval two decades before LLMs. BA Fine Art — information design bridging art and science.
Alignment & people
A decade in the charity sector alongside all of the above — mentoring successive trainee cohorts, presenting to hundreds; two years' residential character training. Responsibility for people, not just systems.