I architect AI systems you can ship. And defend.
I’m an AI architect and ISO/IEC 42001 auditor in Bucharest. I help leaders turn AI ambition into systems they can trust. Three years architecting production AI in legal, banking, and finance, and I still audit against ISO/IEC 42001 today. I design the architecture, ship the system, and know how it will be judged.
| Architected & shipped | Production AI, 3 yrs |
|---|---|
| Audited | ISO 42001 · ISO 27001 |
| Based in | Bucharest, Romania |
AI architecture in regulated, high-stakes domains:
- Legal
- Banking
- Finance
- KYC · AML
- EU AI Act
- ISO 42001
- ISO 27001
- MLOps
The problem
You’re not confused about AI. You’re afraid of the wrong move.
The budget, the board, the regulator, the headline risk. The cost of a bad AI bet isn’t the bill, it’s the regret. What you want in the room is a calm AI architect who has shipped these systems and knows how they fail an audit.
| Architect | Audit |
|---|---|
| I’ve architected the thing.ML infrastructure at 500K+ docs/day, KYC/AML platforms for banks, document intelligence for legal teams. All of it in production, under load. | I’ve audited the thing.I audit AI management systems against ISO/IEC 42001 and 27001, and decide whether they pass. Most people in AI have done one side or the other, not both. |
Services
From unclear AI ambition to architecture you can defend.
| ID | Service | Scope |
|---|---|---|
| S/01 | AI Strategy & Architecture | Where AI earns its money in your business, written so a board can fund it and an engineer can build it.Roadmap / Architecture / Risk / Vendor selection |
| S/02 | Agentic Systems & AI Products | Production architecture for real workflows. LLM apps, RAG, agents, the evaluation that tells you when they drift, and the product around them.LLM apps / RAG / Agents / Evaluation |
| S/03 | AI Governance & Compliance | EU AI Act risk classification and ISO/IEC 42001 readiness, designed in while the system is still cheap to change.EU AI Act / ISO 42001 / ISO 27001 / Policy |
| S/04 | AI Infrastructure & Security | Inference, pipelines, observability, and red-teaming for systems that have to hold up under production load and deliberate attack.Triton / MLOps / Observability / Red team |
Selected work
AI architecture in high-stakes, regulated domains.
| Case | Domain | Detail |
|---|---|---|
| CASE-01 | Finance | High-throughput News Intelligence — 1,000+ articles/secNVIDIA Triton inference pipeline, distributed scraping, extraction, and monitoring. 500K+ docs/day at 99.9% uptime. |
| CASE-02 | Legal | Legal Document AI — RAG in productionPlaybook ingestion, template parsing, and document-aware Q&A for legal workflows where a wrong answer is expensive. |
| CASE-03 | Banking | Compliance SaaS for Banks — KYC/AML · multi-tenantAutomated KYC/AML workflows serving multiple financial institutions, with graph data modeling and full observability. |
| CASE-04 | Assurance | AI Governance Audits — ISO 42001 / 27001I audit AI management systems and information security against the certification requirements, and decide whether they pass. |
Projects
Tools I wrote because I needed them. All of them are open source.
| Project | Stack | Detail |
|---|---|---|
| criv | Rust | Turns a repo’s docs folder into a graph of notes, ADRs, and source references, then fails the check when a reference no longer resolves.REPO · DOCS |
| ste-cli | Go | Checks Markdown and plain text against ASD-STE100, Simplified Technical English. One binary, no dependencies, same verdict every run.REPO · DOCS |
| barrs | Rust | A macOS status bar written against AppKit, configured in Lua, with plugins for CPU, GPU, battery, and the clock.REPO |
| muzik | Python | Downloads, splits, and files music. Wraps slskd, yt-dlp, ffmpeg, and beets, and edits chapters by hand when the sidecar is wrong.REPO |
| xhoard | JavaScript | Archives Twitter/X bookmarks as Markdown. It reads each linked page and files the result by topic.REPO |
| rhoard | TypeScript | Pulls the articles you favorite in Readeck into a folder of Markdown, one file per article.REPO |
Operating model
A practical architecture path you can audit at every step.
The work isn’t model selection. It’s scoping the real decision, facing what the data actually looks like, designing the system, hardening it for production, and proving it’s governed.
- Scope & de-risk — Define the decision the system must improve, and the failure cases you can’t afford. No model talk yet. Stakes first.
- Architect the slice — Design and ship a narrow version on real data, with measurable outputs and a clear evaluation. Proof beats slideware.
- Harden & govern — Add monitoring, latency control, security, and operating controls, then the documentation the EU AI Act and ISO 42001 ask for.
Track record
| NOW | Member of Technical Staff, Footprints AI |
| NOW | CTO, Pythia Sociodynamical Technologies |
| NOW | CTO, Flow Copilot |
| NOW | Auditor, Cert Sud |
| 2026 | ISO 9001:2015 Auditor/Lead Auditor (CQI/IRCA Certified Course) — RINA |
| 2026 | ISO/IEC 42001:2023 Internal Auditor — TÜV Thüringen Karpat |
| 2025 | AI Engineer, CogniSync |
| 2023 | ISO/IEC 27001:2022 External Auditor — CERTINSPECT Register |
Stack
- Tudor Andrei Dumitrascu
- AI/ML
- LangChain
- Pydantic AI
- Claude SDK
- DSPy
- PyTorch
- Scikit-Learn
- NVIDIA Triton
- NLP
- CV
- DL
- Backend
- Python
- FastAPI
- NestJS
- Litestar
- RESTful APIs
- Data
- Pandas
- MongoDB
- PostgreSQL
- Redis
- RabbitMQ
- SurrealDB
- Frontend
- Typescript
- Next.js
- React
- Remix
- TailwindCSS
- Cloud/DevOps
- Docker
- AWS
- Azure
- CI/CD
- Prometheus
- Grafana
- Nix
- Languages
- Romanian
- English
- German
- AI/ML
Start here
Turn the AI decision into architecture you can trust.
Bring a decision, a bottleneck, or a system you need to ship and defend. Thirty minutes, no slides. We’ll know quickly whether I can help.