AI solution architect and research engineer. I work on agentic systems, formal methods, and the mathematical semantics of learned representations, with Bitcoin (UTXO / Script) as my anchored application domain for verifiable coordination.
Themes: geometric and relational semantics for LLMs (compositionality, limits, and obstructions in high-dimensional learned systems, via algebraic topology and differential geometry) • agent coordination: MCP-driven tool use, multi-agent orchestration, trust boundaries • policy-as-code: turning regulation and protocol rules into executable, auditable artifacts • closing the semantic gap between off-chain intent and on-chain truth.
Why the pairing: LLMs and Bitcoin transactions sit at opposite ends of the same question. One is an open, high-dimensional system where structure must be discovered; the other is a rigid, auditable state-transition system where structure is explicit. Both benefit from invariant-centric explanations. I study each with the other in view.
Code The Law — a fully automated agentic pipeline that translates articles of EU financial regulation (MiCA) into executable proof-of-concept applications. One prototype per day through a complete CI/CD chain: generation, review, deployment, publication. A working demonstration of policy-as-code applied to regulation.
BTSL — Bitcoin Transaction Schema Language — a declarative policy-as-code validation layer on top of BIP174/BIP370 (PSBT): schemas describe expected structure, calc + ASSERT encode economic invariants, external checkers get an explicit predicate set for zero-trust replay.
Release: v1.0.0 — Reference Specification [FINAL] · Live explorer: BTSL Playground
| Focus | Link |
|---|---|
| Applied AI writing: tools in LLMs, a taxonomy of agentic workflows, Intelligere (essay on agentic AI) | |
| Design note: a domain-specialized LLM (corpus construction, open-weights fine-tuning, tool layer, evaluation) | DelvingBitcoin |
| Formal semantics for metaprotocols (semantic-gap framing) | formal-metaprotocol-semantics |
| Curated research notes (socio-technical layer, permanence, censorship resistance) | bitcoin_research |
Two agentic systems in production, embodying two orchestration philosophies:
- Code The Law pipeline: a daily autonomous routine driven by declarative instruction files (BOOT.md), decomposed into dedicated tasks through to deployment. Instruction-as-code orchestration.
- p2p-mom (Prompt2Production): a Slack-driven blueprint runner built on a fork of pi, Mario Zechner's open-source coding agent. Inspired by Stripe's internal agent workflows, with deliberately higher autonomy calibrated to a lower risk profile: one run takes a blueprint through to a complete project, typically a PR plus a Vercel deployment. Private for now.
Broader experimentation with Model Context Protocol (MCP) tool use, multi-agent setups, and sovereign agent design. Research questions: coordination surfaces, trust boundaries, autonomy calibration against risk profile, and anchoring assertions in verifiable state rather than mutable off-chain records.
- Education: Mathematics & Applications (Master, research & teaching track) — AMU • Computer Science (M1) — ULB
- Industry: independent AI advisory practice (architecture, agentic pipelines, LLMOps) • CTO (Tokenomeme, infra / GCP) • blockchain dev — Juneo
- Community: speaker & workshop lead — DAO Brussels, UrLab (ULB hackerspace) • CIVIS ambassador (AMU)
AI & agents
Core dev
Formal methods
Bitcoin / PSBT
Automated systems, whether trading agents, coding agents, or regulatory pipelines, only stay honest if their semantics are explicit and their artifacts verifiable. I build schemas, proofs, and static pipelines so that what humans intend and what machines execute remain provably aligned. Bitcoin's consensus layer is my reference case: a long-horizon anchor where this discipline is not optional.
Open to collaboration and critique on the mathematical semantics of LLMs, agent coordination, policy-as-code, Bitcoin formalization, and PSBT / BTSL.