Scattered official data — endpoint behaviour, regulatory registers — turned into clean, sourced, comparable layers that answer engines and agents can retrieve and call.
In most domains the raw inputs already exist — registers, filings, circulars, live endpoints. What's missing is the layer that makes them decidable: parsed, normalised, comparable, current.
Map each source's local vocabulary into canonical structures so data become comparable. Give every datum its source and an as-of date. Capture on a clock, content-hash every fetch, and never overwrite a reading silently.
Websites people can search, answer pages built for how models retrieve information, MCP servers agents call directly. The data updates; the surfaces regenerate.
The method above, run twice — as owned assets.
African fintech regulation
The operating-reality layer for African fintech regulation: who's licensed, what licence is needed, what it takes to hold one, and what may cross a border.
The register normalises 42+ country markets from 156+ official sources, re-checked on a weekly schedule. Every datum carries its source, an as-of date and a link; every page is a deterministic projection of the dataset — a free web register, answer pages, and an MCP server agents can call directly.
Agent-payment rails
The arms-length health, price and history layer for the x402 agent-payment rail. It probes known endpoints on a clock and records what each one actually did — quoted, blocked, unreachable, or simply not observed — and keeps every observation with the raw capture behind it.
Built to be the accountable one in a field of anonymous scores: a published, versioned rubric; a stated correction and dispute process; explicit non-affiliation; and “blocked” never quietly recorded as “dead”.
probe402.com
Citation is won at the passage level, gated by retrieval, forced by uniqueness.
LLMs don't cite websites — they cite passages they retrieved. Winning that citation is Answer Engine Optimization (AEO), and it happens at three gates.
When AEO isn't working, the job is to name which gate is failing — not to treat the engine as a black box. And because citation is stochastic, it's measured honestly: as a rate over a fixed panel of questions across repeated trials, with the candidate set logged alongside the cited set so the failing gate shows itself.
This is how Afriset's answer surface was built. It's the same AEO work, done for clients who need to be findable by machines — not only by people.
One method, written down as a playbook, run the same on every domain.
A domain earns the build only if official reality is scattered, the gap is ownerless, and real decisions are stuck on it. Every candidate runs the same funnel — graded dimensions, explicit blockers, a go/no-go — and most die there, on purpose.
The record before the product. Sources map onto a canonical structure; capture runs on a clock; nothing is published without passing its gates — including honesty about what a source didn't say.
Surfaces are projections of the dataset — registers, answer pages, MCP servers — built for how machines retrieve and cite. Distribution is measured, not assumed. The operator stays on the record.
Capturing what a source says is easy. The work is recording what it doesn't — the missing reading, the endpoint gone quiet, the absence that means something.
Wealth management
Ingests live portfolio holdings and generates investment rationale in the register of the firm's Portfolio Manager — not a template, a system carrying the individual's portfolio-construction logic.
Township labour markets
Designed around how informal workers build trust and work. Full stack: Claude-powered intent classifier, matching engine, React dashboard, POPIA-compliant data architecture with HMAC-tokenised PII.
Senior executive archive
A full local fine-tuning pipeline — transcription, speaker diarisation, LoRA / QLoRA — over years of the subject's interviews, speeches and recorded thinking. Drafts short-form and long-form in their register, trained on considered output rather than surface patterns.
Personal knowledge infrastructure
A routing orchestrator over specialised sub-agents handling research, web-clipping, synthesis, and vault filing — running as a private second-brain system. Telegram interface, live on local hardware, processing and filing continuously.
Pairing fifteen years in finance and strategy with hands-on AI engineering. From a global investment bank to building businesses across sub-Saharan Africa — each experience layer reinforcing the previous and strengthening the foundation for the next. Now researching, building and shipping data sources for LLMs and agents to read.
Engineers struggle with business context; strategists don't ship. We do both.
Tell us what you keep deciding from scattered data — or what machines should be citing you for.
A short note is enough to start.