Dr Maciej Cieslak
AI & Agentic Engineer
Staff Engineer building AI-native software delivery – autonomous agents that raise, test and review production code – and the retrieval systems underneath them. Ten years before that in physics-based simulation and vehicle dynamics, which is where I learned that the problems worth taking on are the ones where the engineering actually has to be right.
Years shipping software
Lines in production
Faster ride simulation
Engineer output with agents
What I do
AI-Native Delivery
Software delivery where autonomous agents raise pull requests, automated QA agents verify them, and a spec-driven workflow governs the lifecycle – with Claude Code as the orchestration layer.
RAG & Retrieval
Production hybrid retrieval (vector + BM25), Azure Document Intelligence ingestion, PostgreSQL with Apache AGE graph queries, and evaluation pipelines to keep answers grounded.
Full-Stack Software
Python/FastAPI, TypeScript/React/Next.js and Angular, REST APIs backed by Azure, AWS and GCP – from proof-of-concept through to production platforms at 500k–1M+ line scale.
Physics-Based Modelling
Multi-domain simulation engines covering mechanical, thermal and electrical systems; Simulink/Simscape CAE tooling, combinatorial architecture search, and vehicle dynamics from model to proving ground.
Signal Processing & ML
ANN-based sensor data modelling, biometric and vibration signal analysis, and ADAS sensor validation – PhD-level work applied from research to production.
Technical Leadership
Led cross-functional teams of up to 11 engineers – defined software strategy, introduced CI/CD culture from scratch, and translated ambiguous problems into well-architected products.
Try it · Agent fleet
A live discrete-event model of an AI-native delivery pipeline – spec, coding agent, automated QA, human review, merge – with rework loops and escaped defects, stepped a simulated minute at a time in your browser. It has two scarce resources, and everything interesting follows from the asymmetry between them: agents run around the clock, while human review exists for one eight-hour shift a day.
Push the fleet past the review gate and throughput does not stall, which is the naive prediction. Something quieter happens instead: review is pinned at a fixed number of pull requests a day, so every extra agent can only add merges through the unreviewed path. The merge mix shifts toward code no human read and the escape rate climbs with it – you buy throughput with quality without ever deciding to. Test coverage is the strongest lever in the model, because it is the only one that buys review bandwidth back.
A model, not a measurement – the rate constants are my own estimates, and the backlog is assumed never to run dry, so it describes delivery capacity rather than demand. The shape of the curves is the claim; the absolute numbers are not.
Try it · Ride dynamics
A live quarter-car + seat ride model – 3 degrees of freedom (wheel, vehicle body, and seated occupant), the same vertical-dynamics physics behind my Fast Ride Model (FRM) work, solved in real time in your browser with an RK4 integrator. The plot reports the acceleration measured at the seat/occupant interface – the ISO 2631 ride-comfort metric. Drag the sliders to feel how the suspension and seat trade off against what the passenger actually feels.
Contact
I'm always interested in hard problems at the intersection of AI, software and engineering – agentic systems, retrieval architectures, physics-based modelling, or bringing rigorous software practice to technically complex domains. Happy to talk.