Closing the Gap Between
Research and Production.

Research signal: Distributed computing, columnar storage, stream processing

Enterprise data platform with reusable components and near real-time serving, delivered while Big Data remained exotic in French industrial groups.

Research signal: Reproducible pipelines, feature stores, model governance, platform engineering

Cloud-native IDP replacing a shadow IT estate. Delivery lead time ÷9. Governance restructured as a service after a prescriptive first attempt stalled adoption.

Research signal: Multi-agent coordination, retrieval engineering, evaluation science, neuro-symbolic reasoning

Evaluation harnesses as infrastructure, deterministic agent orchestration, hybrid retrieval with reranking, inference economics, regulatory alignment under the EU AI Act.

Deterministic state machines over free-form LLM routing

Agent reliability is a property of the harness, not the model.

Framework abstractions that hide the state machine feel elegant until an agent hallucinates a tool signature or loops silently against a token budget. An explicit graph makes every transition typed, observable, and restartable from checkpoint. Termination criteria and step budgets are architectural requirements, not tuning parameters. The trade is developer ergonomics for operational resilience — in production that exchange is never close.

Evaluation harness as infrastructure, not as a gate

Unobservable systems are unoperable. Non-determinism does not exempt.

Classical accuracy metrics do not transfer to generative systems, and distribution shift arrives as semantic drift rather than statistical drift. The functioning pattern: golden datasets versioned alongside prompts, regression suites in CI on every prompt or model change, LLM-as-judge calibrated against human annotation with inter-rater agreement tracked to catch judge drift. Retrieval quality measured separately from generation quality — conflating the two hides which stage actually failed.

Hybrid retrieval and reranking over naive vector search

Reference RAG implementations do not survive adversarial corpora.

Dense embeddings alone underperform on corpora where documents are lexically near-identical but semantically distinct — regulatory text being the canonical case. Hybrid retrieval reconciling BM25 with dense vectors through reciprocal rank fusion, followed by cross-encoder reranking on the candidate set, recovers the precision that pure vector search discards. Chunking strategy is determined empirically per corpus; assumed fixed windows are a reliable source of silent quality loss.

Cost per resolved task over cost per token

Token-level accounting reports regressions as improvements.

An architecture that halves token consumption while doubling retry rates will appear as a 50% cost win under per-token measurement. The correct unit is cost per successfully resolved task. Model routing and cascades handle the tractable majority on smaller models with escalation reserved for genuine complexity; semantic caching absorbs high-repetition query classes. Cost attribution belongs in the same observability plane as latency and error rate, not in a separate finance report.

Selective feature store adoption

Feature stores earn their operational cost in roughly 40% of cases.

Feature stores solve training-serving skew, a real and expensive problem, at the price of substantial operational overhead. The break-even sits at sub-hour refresh cycles combined with genuine cross-team feature reuse. Below that threshold, versioned feature pipelines with enforced schema contracts are simpler, cheaper, and materially more debuggable.

Agentic Systems

Generative AI Engineering

LLMOps & Evaluation

MLOps

Cloud Native & Platform

Governance & Observability