Software engineer — full-stack and AI platforms. I ship systems end to end, and I measure before I claim. Based in Paris, available for full-time roles from February 2027.

Nothing below is merged yet; each line says where it stands.

• apache/iceberg-python#4003 — issue, open. Above 200 values in an IN predicate, PyIceberg turns off file pruning: a delete then plans 20 of 20 data files instead of 1, rewrites the same single file, and reads the other 19 for nothing. One key more, twenty times the I/O. Shipped with a 40-line reproducer, a 240-run benchmark with medians and dispersion, a control experiment that rules out predicate size, and the 2020 Java commit that introduced the guard.
• apache/iceberg-python#4004 — PR, under review. In / NotIn built their literal set twice: 200,000 objects for 100,000 distinct values. Predicate construction is 42% faster at 100,000 values. Reviewed by a maintainer, revised, and the revision also fixes a silent bug where a predicate built from a generator came out empty — a filter that matched nothing, with no error.
• huggingface/huggingface.js#2510 — PR, open. A @huggingface/hub test assumed the platform supported symlinks instead of detecting the capability. It failed on Windows and was invisible to an Ubuntu-only CI.
• NVIDIA/garak#2217 — issue, open, independently reproduced. English-only detectors misjudge a target that answers in another language, in both directions: a French refusal scores as a successful attack, and French toxic content scores clean. False positives are visible; false negatives are not. Discussion with the maintainers is open on whether this belongs in core or behind a flag.

• AFTERSHOCK — real-time seismic lakehouse. USGS events through an idempotent Kafka (KRaft) producer into Spark Structured Streaming, Delta tables on HDFS in bronze/silver/gold, four Airflow DAGs chained by Datasets, replay after a hard kill proven to write no duplicates. It measures something nobody archives: of 2,128 earthquakes tracked across their full version history, 735 (34.5%) had their magnitude revised after publication — the number that triggered the alert was not the final one.
• Taintrace — sanctions screening over a Bitcoin contamination graph. OFAC and OpenSanctions ingested by Airflow as streams without disk writes, Bitcoin screened over Kafka, taint propagated over n hops in Neo4j, the whole chain running as containerized services that start with one command. Traced 512.25 BTC received by an address on no sanctions list, 11.07% tainted at three hops — the case regulators care about and a single-hop check misses.
• Phantera — LLM outbound platform, solo-built, pre-launch (private). Durable orchestration on Inngest with deduplication by event key in Redis, Stripe usage-based billing, and a CI gate built on an LLM evaluation harness (F1, MAE, Spearman) with NVIDIA garak red-teaming and prompt-injection detection.

• I benchmark with repetitions, medians and dispersion, never a single run — and I throw away a series when the control says the machine drifted. One of the Iceberg series went in the bin for exactly that reason.
• I write down what I did not prove. The Iceberg issue says plainly that I could not reproduce the reporter's 1054 s, and that the fix does nothing for uniformly spread keys.
• I kill my own ideas when the code disagrees. A second optimisation looked like a 4x win until I read Literal.to closely: it bounds values rather than converting them, so the shortcut would have made pruning wrong. It is out.

Python · TypeScript · Java · event-driven microservices (Fastify, NestJS, Spring Cloud, API gateways, Keycloak) · Kafka · Spark Structured Streaming · Airflow · Delta Lake · Apache Iceberg · Neo4j · PostgreSQL · MongoDB · Redis · Next.js · React · Docker · Playwright · LLM evaluation and red-teaming

Portfolio · LinkedIn