AI Engineer · Hoxton, London
You build the best spreadsheet modelling agent in the world. This page is a model review of the candidate, run in your own grammar: eight findings graded by severity, four scenarios flexed, two of them failing.
Every claim below traces to a public repository, a shipped product, or a verification link.
Delivered for
Harness · tools · context
The harness runs on a loop. A pattern I have explained twice becomes a skill. A task I repeat every session becomes a hook. Each pass makes the next one cheaper.
Agents were spending their context window searching a repo they had already read. I wired graphify, an open-source code indexer, into two hooks: one catches grep, rg, find and fd before they run and points the agent at the graph instead, and one rebuilds that graph when a session starts, so it stays current on its own.
Eight of the skills are scoped to this repository. They build these pages: five subagents in two waves against a shared contract file, each owning its own files so they can run at the same time. The MCP server I wrote sits over the same function library the web UI and the CLI call, so an agent picks up the tools I use.
Built to be checked
Proof that travels without the source
An allocator reading hundreds of applications can properly check only a handful, and a founder whose best work sits in a private repository needs a way to show it. takes custody of the build, measures it in a sandbox, and returns a signed verdict that cites every piece of evidence it read. The investor gets a result they can act on, and the founder keeps the code.
Model review
Run the way your reviewer runs one. Four clear on public evidence, two sit at medium, and two are graded High.
Hooks on two events, subagents with a contract file, an MCP server he wrote. A repeated pattern becomes a skill, a repeated task becomes a hook. The posting's second bullet describes his Tuesday.
.claude/settings.json
, this site, and 's GraphQL gateway are all TypeScript. React 19 and TanStack across the front ends.
projects.ts · wiregent
195 commits and roughly 145,000 lines in eight days, and in the same window test files went 17 to 47 while source files went 101 to 222.
projects.ts · wiregent
Design was the professional practice before code, from 2008. Design-system owner on end to end in Figma. Joined as a UX Engineer, not an engineer.
profile.ts · DESIGN_PRACTICE
The team is ex Jane Street, Bloomberg, Cambridge, UCL. He is none of those. What stands in its place is nine years of shipped product across seven countries and a public repository trail anyone can open.
Accounting SaaS, a multi-module ERP with ledger and dues tracking, and a double-entry hash-chained ledger with integer money. Closest to your product: , where QA authors write test logic in a spreadsheet and the platform compiles it to three execution engines. He has never built a three-statement operating model or an LBO.
projects.ts · optimo, elobooks
Go is the closest systems language on record and it runs 's pipeline. This posting lists Rust as a plus. Learning it is the ramp, and ownership and lifetimes are the new part.
skills.ts · no entry
Evaluation pipelines and data collection are real and shipped. Fine-tuning and reinforcement learning are not. Of everything the posting asks for, this is the widest gap.
Systematic
Written out the way your methodology section is: one row per requirement, all ten on the page.
Between 20 and 27 July 2026 went from a scaffold to a working platform: 195 commits, 348 files touched, roughly 145,000 lines added, source files from 101 to 222 and test files from 17 to 47. It sits at 1,937 tests today. On models, I built the layer instead of picking a favourite: three drivers behind one interface, Anthropic and Gemini alongside a deterministic driver that needs no key, no model and no network. Any timeout or malformed response falls back to it instead of erroring the request, so the platform still runs with no API key set at all. Claude leads the delivery work with two Anthropic certifications behind it, and runs Kimi through OpenRouter because that stage filters at volume and does no reasoning.
The harness enforces context optimisation: a PreToolUse hook matches grep, rg, find and fd and returns an instruction to query a scoped knowledge subgraph instead, while a SessionStart hook keeps that graph current. The loop around it is the point: a pattern explained twice becomes a skill, a task repeated every session becomes a hook. Eight of those skills are scoped to this repository, and 5 subagents build in two waves against a shared contract file with strict file ownership. Tools are written to be agent-first: 's web UI, CLI and MCP server are three thin wrappers over one function library, six MCP tools, and the Claude Code harness tested the tool by calling the project's own Skill and MCP server before a human ever opened the web UI. runs the product-scale version, with founder and investor agents negotiating under a per-run token budget, a prompt firewall and metered spend.
The first two halves are shipped. I modelled an eleven-category diligence taxonomy with four weight profiles and grounded it in original research: a 6,080-company outcome study, ten world-market studies of what allocators require, and a moat analysis that flags every unsourced claim. The eval decision that makes it worth anything was to report an unevaluable rule as unproven rather than zero, because collapsing "there is no API" into "the API is not growing" was failing the founders the product exists to find. The prompt firewall has a measured red-team result: 41 planted sentinels probed, 0 leaked. gates every generated script against the scraped source bodies with grounding evals before it reaches synthesis.
SFT and RL are the honest half. No fine-tuning runs, no reward modelling, no RLHF or RLAIF in production. He has built the data collection and the evaluation harness that a training loop consumes and has never closed the loop itself, so a first quarter owning that would be a genuine first and should be priced that way.
Three of ten are partial. All three are on this page before they are in an interview.
Design was the professional practice before code was, running as paid delivery since 2008. Two hundred logos drawn, hundreds of UI components, sixty-plus applications and websites delivered internationally. He is design-system owner on end to end in Figma, and he joined as a UX Engineer, not an engineer. Working with a designer here means arguing in the same vocabulary.
Co-founded in 2017, grew it to 22 engineers, then ran it distributed and remote-first, on a track from UX Engineer through Principal to Engineering Manager to CIO before stepping out of the operating role in August 2026. The part that matters for this line is what stayed behind: alumni now lead at organisations in Norway (2), Canada and Germany, plus two leading local Bangladeshi tech companies. He is currently sole engineer on , so the recent proof is ownership.
The evidence is public and checkable. is live at wiregent.com with 1,937 tests and a repository that publishes a row-by-row inventory of what each capability really does, so a reader can tell shipped behaviour from a demo. ran six years as a Zürich client before Loopcloud acquired it in 2026. 's ERP is still run by the client. 's patient records are still in use at Bangladesh's largest public medical university.
What is not there is the pedigree the posting's own team is described by. No Jane Street, no Bloomberg, no Cambridge or UCL, no competitive-programming ranking and no big-tech badge. If top centile is being read through that lens, this is a clean no. If it is being read through what someone has built and can demonstrate in a take-home, the case is on the table.
All four. Agents: five subagents with strict file ownership and a contract file between them, plus the negotiation harness where founder and investor agents haggle under a token budget, a prompt firewall, metered spend and stop timers, and ’ six-agent hub-and-spoke on Google’s ADK. MCP: a server I wrote exposing six tools over stdio, sharing one function library with a web UI and a CLI. Skills: 82. Hooks: SessionStart and PreToolUse, the second to stop the agent burning context on raw search.
exists because a client's launch nearly stalled on a thousand-photo gallery that needed every face blurred except one, with no time for a manual pass; it shipped overnight as a web app, a CLI and an MCP server. Nine years co-founding and building it to 22 engineers is the long version. The stubborn version is a 900+-day consecutive French streak with A2 grammar explainers published publicly, which is agency applied to something with no professional payoff at all.
Five of the eight are strong and current. Building agents is the core of the last year. TypeScript is the primary language across , this site and 's gateway. React 19 with TanStack across the front ends. Devops runs Docker, Fly.io, Cloudflare Workers and R2, GitHub Actions and multi-target builds.
Competitive programming, trading and big-tech employment stay open here. What stands in their place is nine years of shipped product across seven countries and a public repository trail anyone can open. Rust is the fourth, and coming from Go it is a real ramp.
Five days a week in person is what he wants, and London is where he wants it. The city is Europe’s hub, where frontier AI is both built and used, so one seat reaches more business domains and more unfamiliar problems in a year than anywhere else he could sit. That competition is the appeal. The personal tie is already there: his sister lives in London, and ran Dubai to London in 2024.
The mechanism is a company-sponsored Skilled Worker visa. That is a real cost and a real timeline for a small team, and it is worth raising in the first conversation.
Scenario analysis
Four cases. Two clear today, one ramps, one has not been run.
Scenario analysis
Beyond the role
The posting asks for an engineer who becomes a technical leader and works with a designer. Those are two of the five surfaces I have already owned end to end on the same product.
The mark, and the system
Design delivery has been a paid practice since 2008, largely through 99designs before it was ever a job title. ’s own mark, its brand and its business cards are mine. I own ’s design system end to end in Figma, and the hundreds of UI components behind sixty-plus applications came out of the same hand.
The code, and the ledger
TypeScript first, Go where a pipeline earns it, Python where the sandbox stage lives. is the Go one: 52,937 H3 hexes across London and Paris, each carrying a two-hundred-feature vector, every hypothesis written down and registered before the product was allowed to claim anything from it. Earlier, compiled a Hyundai QA team’s spreadsheets into three execution engines. runs Python gRPC behind a TypeScript GraphQL gateway because that is what each half needed. Underneath sits a double-entry ledger: a sole-writer module that throws on unbalanced legs, a hash chain per account, integer minor units end to end, and balances recomputed from the legs on read.
The room
I co-founded in 2017 and grew it to twenty-two engineers before running it distributed and remote-first. Two alumni now lead teams in Norway, two lead at top local firms, and an early-career direct report from the in-office years is now engineering in Sweden.
Verifiable
Your security page leads with SOC 2, GDPR and ISO 27001 because a finance customer asks before they evaluate. The same question gets asked of an engineer, and this is the answer with links attached.
See for yourself
The posting's first step is a technical interview or a take-home. That is the right filter here. My case is a body of work, and a take-home reads one faster than a CV does.
