What I'm building
Fieldwork
The reference stack
Six instruments along the path a venture actually takes: decide, build, control, approve, revalidate, explain. Each one is open to use on its own, and each one can be applied with support.
- Status: v0.1
Decide step of the lifecycle. Platform Decision Kit
Triage and discovery instrument for platform decisions in existing systems: tagged hypotheses, a discovery brief, a seven-slide skeleton.
runs entirely locally, no data leaves your browser
- Status: v0.1
Build step of the lifecycle. Local Agent Pipeline
A fully local multi-agent pipeline with a tamper-evident audit trail, an action-space whitelist and a human-in-the-loop gate — a reference pattern, not a framework.
- Status: v0.3
Control step of the lifecycle. Agentic AI Governance Toolkit
Lifecycle, risk scoring and control intensities for AI agents in regulated organisations: a runnable evaluator, EU AI Act and DORA checklists, an action authority matrix, and a gate that emits governance findings as a build artifact.
- Status: v0.3
Approve step of the lifecycle. RAG Approval Blueprint
Approval file, control catalogue with test procedures, deletion evidence, a worked-through approval run: the complete approval process, documented.
- Status: live
Revalidate step of the lifecycle. Legal status record
An operable register: does the law still say what your approval cites? 40 EU acts resolved against the primary source, register builder, record as text and JSON.
the selection stays in the browser; on request only CELEX identifiers are queried
- Status: Case
Explain step of the lifecycle. Approval teaching case
One case played through: where the error came from, which architectural gap allowed it, what evidence existed, and how that turned into a go or a no-go.
Embedded
Tools for embedded decisions. Most compute in the browser without a single byte leaving the machine; the Decision Ledger stores a record the moment you save.
Embedded Decision Ledger
Status: v0.1Enter a make-or-buy question and turn it into a decision record: options, evidence with its verification state, who decides and who approves, review triggers. The tool computes the status from the fields. Export as Markdown or JSON.
Run the instrumentDrop in an ARXML file and the record gains an evidence line with a content digest, marked “verified by machine” — the one line in it that does not rest on self-assessmentDelivery delta report
Status: v0.1Drop in two ARXML states — the last delivery and the new one — and read what moved: schema revisions, occupied layers, communication surface, variant count. Every number names the files it came from. Runs entirely in the browser, no schema validation.
Run the instrumentIf two states share no anchor, the tool refuses the delta rather than subtracting two unrelated sumsAI Deployment Ordinance
Status: v0.1Where may agents work in your embedded programme, under which independent check, and who is accountable? Classify work areas — free, with check, proposal only, blocked — and take away four outputs: a team ordinance, a management one-pager, an audit trail and an allow/ask/deny permissions file for your agent configuration.
Run the instrument“Free” is unreachable without an existing independent check — that sits in the construction, not in the proseEmbedded Make-or-buy Kit
Status: v0.1Capture building blocks, variants and constraints; get a reasoned make, buy or hybrid hypothesis per block — with a decision file in text and JSON to take away.
Run the instrumentGerman-language instrument; computes entirely in the browser; prices, variants and names never leave the machineProduct Safe Delivery & Scaling
Status: Version 0.1The rulebook underneath the instruments: a systems architecture council with composition and quorum, an ADR template with a review date, decision rights, the four agent zones, five gates to market, and the licensing limit.
Run the instrumentwritten to be adopted and cut down; every factual claim carries a source number
Curated
Awesome AI Governance
Status: ListA curated list: regulation, standards, risk frameworks, security and open-source tools for governable AI. Focus: agentic systems in regulated environments. Curated means: selected, placed in context, maintained.
Built & shipped
Products I've built, launched, and still run.
razbiram.com
Status: LiveUnderstand Bulgarian, word by word: an AI-assisted learning platform with evaluated language-model prompts (golden sets).
razbiram-anki
Status: BrowserYour own Anki decks in the razbiram style, runs entirely in the browser, no install, no account.
Critique Incognito
Status: iOSCapture restaurant reviews on site: a fixed checklist, every value measured or estimated, a draft with a source for every sentence.
Governance tools under MIT,the curated list under CC0.
Clone them, adapt them, use them in-house. Questions and additions belong in the repo's issues.
See the offering