Built, verified, in use.
One pattern underneath: the inputs a flourishing society runs on — energy, materials, industrial capacity, compute — and the joints between them. One design DNA throughout: provenance on every number, deterministic verification before any model's judgment, honest accounting of what is not verified. Client engagements are described by shape only — sector-vague by intent; the work speaks, the names stay private.
Not products. Working instruments, the way a carpenter owns his bench. They exist because the engagements demanded them.
A governed autonomous build system
Turns a plain-language mission into working, verified software: multi-agent planning, parallel ticket execution, twenty-plus deterministic quality gates before any LLM judgment, and a learning loop that feeds each build's lessons into the next. Roughly 122,000 lines of TypeScript, produced in ten weeks by a non-developer — every version generated by the platform building itself, across sixty-plus self-build iterations.
An event-sourced work memory
Every AI-assisted working session, indexed and searchable: an append-only event spine with per-project row-level security and a sensitivity policy that keeps confidential material on-device — confidentiality enforced at the database tier, not trusted to the app.
A project operating system
The written method the tools serve: destination documents ratified before building, disposable mocks before missions, gates where builder self-reports are never accepted as evidence, provenance and refusal as first-class outputs. Written up as a portable, two-track system — governed, and fully autonomous for fast first cuts.
A large-scale industrial investment, made computable
For a confidential Nordic development: a deterministic feasibility engine spanning process physics to monthly three-statement financials — every quantity carrying source and grade, scenario cases as first-class objects, and a machine that refuses to extrapolate rather than flatter. Built to survive bank-grade diligence, with the audit trail as the product.
Commercial strategy as a falsifiable memo
For a European infrastructure company financing at scale: a go-to-market strategy argued from financing mechanics rather than sales narrative — anchor relationships as instruments of bankability — written as a hypothesis with named assumptions and the observations that would disprove it.
Committee-grade diligence, from public data
For a private-markets investor: an intelligence platform that turns public clinical, regulatory and corporate sources into diligence packages in the firm's own memo format — with every claim source-cited and a catalog of what could not be verified. Taken through to a scoped commercial pilot.
Investment-grade read of an AI platform
For a technology-driven investment firm: a multi-pass audit of their proprietary agent platform — separating shipped capability from architectural intent, verifying the competitive thesis against live sources, and surfacing a material defect with a scoped remediation. Builder's judgment applied with a diligence officer's discipline.
Working software and the strategy above it
For a manufacturing-software company: a production CRM — built, integrated, security- audited — paired with a reviewed product thesis for the AI intelligence layer above it. Software and strategy from the same hands, which is rather the point.
Data-center economics as physics
An investment model treating AI compute as a chain of thermodynamic and supply constraints — a min-of-ceilings solver over a 162-equation specification, sixteen scenarios, source equations separated from interpretation. Fed by a companion market- intelligence pipeline with an LLM-processed news feed over a verified research base.
Clinical-trial intelligence, compliance-first
A strategic layer for trial operations built on the industry's own data standards, with a hash-chained audit ledger and pseudonymization at the AI boundary — regulation treated as load-bearing architecture, not a feature checkbox. Alongside it, a verification-first pharma pipeline whose signature output is the list of what public data cannot prove.
A self-verifying quantum simulator
A tensor-network circuit simulator that checks its approximate engine against exact contraction and analytic physics targets — thirty-five passing benchmarks, explicit about every convention and every claim it does not make. Built to learn the territory properly.
Everything here ships with test suites, audit trails and stated gaps — a habit formed over twenty years around credit committees, where claims that can't survive checking don't survive at all.