Mathematician by training. I build AI systems for enterprise IT that has to work on Monday.
|
I own an AI platform as an architect: agentic environments for SAFe value streams with dozens of teams — the workflows, the MCP servers, the guardrails. These repositories are where the ideas get tried out before they have to hold up in a place where an outage is a headline. Twenty-five years of owning quality in critical systems left me with one bias: what you don't measure doesn't work. It shows in almost everything below. $ oliver --now
role AI platform architect · product owner
scale SAFe value stream, dozens of teams, compliant
building agent environments · MCP servers · Tools · RAG ·
studying math, math with AI, distilled by machines, a lot of IT
playing helpmates, chess exhaustively, math puzzles |
The trapped knight. Always jump to the lowest unvisited number on a square spiral. After 2016 squares, every reachable neighbour has been used — stuck on 2084. Colour runs from the first move to the last. |
How do you move a legacy system to a modern architecture when the business logic was never documented? In 2025 we did exactly that — 50,000 lines, with agents reconstructing the domain rules from the code itself.
flowchart LR
L[("COBOL · Smalltalk<br/>Fortran · Pascal")] --> D
subgraph W ["LangGraph workflow, every step an MCP server"]
direction LR
D["Discovery<br/><sub>file catalogue</sub>"] --> A["AST analysis<br/><sub>tree-sitter</sub>"]
A --> G["Dependency map<br/><sub>graph + cycles</sub>"]
G --> U["User stories<br/><sub>LLM, with confidence</sub>"]
end
U --> B["Migration backlog<br/><sub>readable by stakeholders</sub>"]
style L fill:#0b4f6c,stroke:#0b4f6c,color:#fff
style B fill:#1b9aaa,stroke:#1b9aaa,color:#fff
Most reverse-engineering tools stop at syntax. The interesting part is recovering business intent — and then admitting how confident the machine actually is.
| repo | what it does |
|---|---|
| archaeocode | Automated software archaeology. Point it at a legacy codebase, get user stories with acceptance criteria and confidence scores, plus a dependency map. COBOL, Fortran, Pascal, Smalltalk alongside Java, Python, JS, TS. |
The model is rarely the problem. The context is. Each of these stands alone — together they are one agent environment.
flowchart LR
H["AI host<br/><sub>Claude Code · Codex · Copilot · Kiro</sub>"]
S["ctx.skills<br/><sub>time · ontology · rlm</sub>"]
Z["aegis-mcp<br/><sub>default-deny · namespaced · audited</sub>"]
R["repo-ctx<br/><sub>symbols, dep graphs, 12 languages</sub>"]
M["semrag<br/><sub>Qdrant + Neo4j hybrid</sub>"]
K["knowledge-mcp<br/><sub>document RAG</sub>"]
Q["SAIRmcp<br/><sub>challenge API</sub>"]
S -.-> H
H <==>|"MCP / stdio"| Z
Z --> R
Z --> M
Z --> K
Z --> Q
style Z fill:#d1495b,stroke:#d1495b,color:#fff
style H fill:#0b4f6c,stroke:#0b4f6c,color:#fff
| repo | what it does | |
|---|---|---|
| repo-ctx | MCP server giving an assistant searchable access to any Git repo: real symbol extraction, dependency graphs, architecture analysis across 12+ languages. Local, GitHub, GitLab, whole orgs. |  |
| ctx.skills | Agent skills for what LLMs are notoriously bad at: telling the time, keeping knowledge structured, reading inputs larger than the context window. Stdlib-only Python. | |
| semrag | Local-first enterprise RAG: unstructured vector similarity (Qdrant) blended with structured graph intelligence (Neo4j) in a stateful LangGraph workflow. | |
| knowledge-mcp | Three MCP servers for document ingestion and semantic search over enterprise document stores. |
An agent you cannot constrain is not a tool, it is a risk. An answer you cannot check is not an answer.
| repo | what it does | |
|---|---|---|
| aegis-mcp | A zero-trust gateway for MCP, in one Go binary. Capability profiles are default-deny, every tool is namespaced by its origin server so it cannot be shadowed, escalations go through human approval, every decision is audited. Fails closed. | |
| urllm | Point it at any URL, get a grounded GDPR and security audit. Deterministic extraction of the full fingerprint first — cookies, CSP, third-party domains, PII forms, trackers — then the LLM. No raw HTML dumped into a prompt. --fail-on gates your CI. |
|
| SpecBench | Benchmarked survey of spec-driven development frameworks — the counter-movement to writing code by hoping. |
My training, and still the thing I read for pleasure. Currently it is where AI and mathematics meet: I am competing in the Mathematics Distillation Challenge of the SAIR Foundation.
A magma is a set with one binary operation and no laws at all. Impose a single identity, say
and you have picked one of 4694 equational laws. Which of them imply which others? That is
the question the Equational Theories Project set out to answer with automated provers and human proofs together.
| repo | what it does |
|---|---|
| magmaexplorer | An interactive REPL for the human half of that work. Six mechanically-verified derivation primitives; an LLM may propose derivation steps, but the tool re-executes and verifies every one before it is accepted. Exports to YAML and Lean. |
| SAIRmcp | MCP server for the SAIR Challenges API, so an agent can work the competition with me. |
| the-trapped-knight | The picture at the top, generalised to |
How far does a greedy knight get? — computed with the library
| knight | path length | x-diameter | y-diameter | density |
|---|---|---|---|---|
| 2016 | 55 | 56 | 0.65455 | |
| 3723 | 92 | 91 | 0.44470 | |
| 4634 | 81 | 81 | 0.70629 | |
| 13103 | 125 | 125 | 0.83859 | |
| 27154 | 277 | 278 | 0.35262 | |
| 27430 | 277 | 283 | 0.34991 |
Density is
I love playing chess with a decent FIDE rating And chess is sometimes where the mathematics and IT gets a board. In a helpmate, both sides cooperate: Black moves first and helps White mate in the fewest moves. Because both sides want the same thing, it is a cooperative shortest-path problem rather than a min/max tree — so whole material classes become tractable, not just single positions.
| pieces | classes | status |
|---|---|---|
| 2–4 | 66 | ✅ complete |
| 5 | 220 | ✅ complete |
| 6 | 645 | 🚧 9 done, 636 to go |
| 7+ | — | ❌ needs an out-of-core generator that does not exist |
295 tables, 38.5 GiB block-compressed, on Hugging Face. The deepest mate in the corpus is h#17. For every legal position the table stores not just the distance to mate but how many optimal solutions tie — so it can tell you whether a composition is sound or has a dual. Anyone can search for a mate; this answers uniqueness, for every position at once.
286 of the missing six-piece tables need only 32 GiB of RAM and about a day of CPU each. If you have a machine that idles overnight, you can compute something nobody ever has. → How to contribute
| repo | what it does | |
|---|---|---|
| helpmate-tablebase | Every helpmate in a material class, solved exhaustively. Not a solver you point at one position — a table that already knows. |  |
| ChessMG | The move generator underneath: magic bitboards in C++20, zero-copy Cython bindings, 200M+ moves/sec — roughly 50× a pure-Python generator. Verified against perft to depth 8. pip install chessmg. |
 |
Everything else that is public
| repo | |
|---|---|
| amazons | Game of the Amazons — engine and analysis |
| notify-mcp | MCP server for team communication and notification |
| the-trapped-knight.js | The spiral, animated on a canvas |
| CollatzViz · Collatz | Collatz sequences, visualised and computed for arbitrarily large integers |
| collision | A small neural network for collision detection |
| portfolio | Stock portfolio proof of concept, Streamlit + yfinance |
Daily driver: Claude Code · Codex · GitHub Copilot · AWS Kiro · LinkedIn · Hugging Face ·