Reliable, measured agentic systems in production.
Four engagement areas, as a mission or a turnkey deliverable. Each one is backed by systems actually built: the details can be verified in the projects and the track record.
Multi-agent architectures that hold up beyond the demo: Plan-and-Execute orchestration of ReAct sub-agents with LangGraph, human-in-the-loop safeguards, MCP tooling, SSE streaming and prompt caching to keep cost and latency in check across multi-turn conversation. The kind of system I carried to production as founding engineer at a pre-seed AI startup.
See it in practiceMulti-agent architectureUGC video engine
Multimodal ingestion pipelines at scale: OCR, image-to-text, semantic chunking, embeddings and hybrid search, on Databricks as well as on an open source stack. With the experience of having written every stage by hand, from chunking to conversational memory, before frameworks made them standard.
See it in practiceProduction multimodal RAG pipelineQuantum Insights, a RAG pipeline
What separates a prototype from an operable system: evaluation frameworks with Ragas and LangSmith, LLM-as-Judge, tracing and cost observability. Put into practice across five candidate RAG pipelines, where the in-house pipeline outperformed the commercial baseline on both precision and recall (Precision@k 0.82 vs 0.68).
See it in practiceFive-pipeline RAG benchmarkAgent trajectory evaluation
Getting systems to production and keeping them there: AWS Bedrock and open source endpoints, FastAPI, Docker, CI/CD, MLflow for model versioning, and the feedback loops that let teams iterate on data rather than intuition.
See it in practiceMeasured prompt caching and prompt-as-codeIndustrialized credit scoring
As a certified Lean Six Sigma Green Belt, I bring production-line discipline to AI systems: metrics defined before the solution, results read critically, decisions settled on numbers.
Used to distributed teams and distant timezones, working directly with a founder or product owner. On long missions as well as turnkey deliverables.
Versioned, documented, transferable code: conventions, CI/CD and observability in place at handover. A system the team still controls after I leave.
Browse in the portfolio
Models industrialized end to end: credit scoring traced across 288 runs, vision under degraded conditions, survival analysis on a 50-robot fleet.
Real-time pipelines at scale: in-stream computation on Kafka and NiFi, multimodal ingestion on Databricks, 22.7 million on-chain transactions consolidated.
AI wired into production: a conversational agent over live telemetry, a maintenance assistant over scanned OEM documentation.
Before AI, measurement: the carbon footprint of a plastics plant quantified at 2,444 tCO2e, predictive maintenance and TPM on a robot fleet.
Tell me the context and the goal, and I will tell you how I would approach it, as a mission or a turnkey deliverable.
EN
FR