Work experience
I've been working as a developer since 2011, at nine companies. Some of the projects are described on the projects page.
I lead the data team technically and I'm responsible for the data platform end to end: its architecture, a good part of its implementation, and helping teams use it. I make the architecture calls (Trino and Iceberg, VEGA, replacing Qwen with GLiNER) and work with product teams to solve their performance and feature problems.
- Replaced Qwen2.5-32B on GPU with a fine-tuned GLiNER service on CPU at an order of magnitude lower cost. Competitive with Qwen overall, behind mainly on brands. Measured against judge-labelled and human-calibrated references, then run in shadow mode on live traffic before the switch. →
- Designed VEGA, the asynchronous query API product teams read the lake through, retiring at least six per-team data copies, several ETLs and a costly Spark process. →
- Introduced Trino as the lake's query engine and Apache Iceberg as its table format, replacing Delta Lake, with Trino taking over most of Spark's query work. I maintain the 15 TB+ lake's Iceberg tables, which 20+ streaming consumers read and write, and finished the move of enrichment from batch to real time. Geolocation lookups now run straight on the lake through Trino, and we decommissioned the MySQL copy we kept only for performance. →
- Built the interest-classification pipeline (LoRA fine-tuning, cross-encoder reranking, a model-agreement rule, vLLM on Kubernetes), taking human-judged strict precision from 36% to 52%. →
- Cut Trino query failures from 2.2% at worst to under 0.2%, and internal network per query by 93%. →
- Built tools other teams use: an agent skill (a plain skill file that works in Claude Code and Codex) that lets them write and improve their own VEGA queries and build reports faster, and a Slack bot that alerts us when a data-lake table degrades.
- Ran internal sessions on working with agents (multi-agent setups, context management) and on fine-tuning with LoRA and reinforcement learning, and mentored colleagues on adapting decision models (small models that answer a yes/no question with a score) to our domain, from building the training and test datasets to tuning the model.
Designed and built a data platform for fraud analysis in motor insurance claims, and took it to production. Scala, Kafka, Akka, Cassandra.
Batch and streaming systems for clients; research on real-time processing architectures. Built ML model deployment pipelines on Kafka Streams, with monitoring on the ELK stack.
Node.js/Express APIs, OAuth2 for internal and third-party authentication, and planning of the data-analysis stack (Elasticsearch, Spark, Kafka).
SmartData: real-time footfall counting from WiFi device detection with Spark Streaming and Cassandra. Also IoT monitoring over WebSocket and end-to-end encrypted live video streaming.
Medical research data-collection platform with IoT biometric device integration.
Stack
What I work with
Currently, grouped by area. Technologies from earlier roles are listed at the end.