Experience#

Machine learning and forecasting

Time series and demand forecasting, cost and schedule models, gradient boosting, survival and risk scoring, portfolio optimization, computer vision, conformal prediction and quantile methods for calibrated uncertainty, and explainability that shows which drivers move each forecast or score. scikit-learn, XGBoost, LightGBM, statsmodels, SciPy, PyTorch, TensorFlow, SHAP, MAPIE, skfolio, OpenCV.

LLMs and AI agents

Agentic workflows with tool use, retrieval-augmented generation, structured extraction, classification, text-to-SQL and local models. Pydantic AI, Claude Agent SDK, LangChain, LangGraph, LlamaIndex, LiteLLM, MCP, Transformers, sentence-transformers, MarkItDown, Ollama. Anthropic, OpenAI and Gemini models via their APIs, Bedrock, Vertex AI and Azure OpenAI.

Speech, language and localization

Speech-to-text across providers and languages, speaker diarization, word and diarization error rates against human-annotated transcripts, text normalization, and checks that localized output stays in the user's language. AWS Transcribe, AssemblyAI and Azure Speech for transcription, jiwer for scoring, ffmpeg for audio preparation, spaCy and NLTK for text processing.

Evaluation and MLOps

Evaluation designed before the build: ground-truth datasets labeled together with domain experts, baselines a model has to beat, leave-one-out and cross-validated backtests, LLM evaluation harnesses, data validation, experiment tracking and tracing of runs, so that every change to a model or prompt is measured against the same benchmark. Pydantic Evals, Langfuse, Logfire, MLflow, OpenTelemetry, pandera.

Data engineering

Batch and streaming pipelines, warehouse modeling, data quality checks, query tuning and SQL at scale, from a laptop prototype to the production warehouse. DuckDB, Polars, Pandas, PyArrow, Parquet, Spark, Dask, sqlglot, SQLAlchemy, SQLModel, Alembic. PostgreSQL, PostGIS, MySQL, MongoDB, Redis, Neo4j, Elasticsearch, OpenSearch. BigQuery, Snowflake, dbt, Dataform, Airflow, Celery, Kafka.

Data analytics

Dashboards and interactive apps that people use after the project ends: Looker, Looker Studio, Metabase, Power BI, Streamlit, Dash, Plotly, D3, JupyterLab, marimo, Matplotlib, Seaborn. Technical reports, white papers, handover decks and executive summaries that non-specialists can act on, written with LaTeX, Zensical, MkDocs Material or python-pptx and co-authored with the client where needed.

Cloud, infrastructure and operations

Reproducible infrastructure on AWS, Google Cloud and Azure, from managed services to a hardened Linux server, with secrets, backups, monitoring and alerting in place, and deployments that run from CI rather than by hand, with every change reviewed and versioned in Git. Terraform, Docker, systemd, lighttpd, fail2ban, ufw, Let's Encrypt, Fabric, GitHub Actions, GitLab CI, Grafana, MinIO, Cloudflare.

Geospatial, mobility and simulation

Movement analytics from cellular, WiFi and fare-card data, origin-destination matrices for planners, traffic and transit simulation, spatial joins, network analysis and interactive maps at city scale, with aggregation that keeps every individual anonymous and the results compliant with GDPR from the first query. GeoPandas, Shapely, pyproj, Rasterio, xarray, PostGIS, NetworkX, SUMO, Folium, Leaflet, deck.gl.

Software engineering

Python first: uv, Pydantic, pytest, Hypothesis, ruff, pylint, and ty and Mypy for type checking. Rust with PyO3 and maturin for speed, plus SQL, Bash, Make, TypeScript with Node.js, vitest, TypeBox and ESLint, Lua, Java, Kotlin, Scala, C/C++ and Cypher. APIs and CLIs with FastAPI, Flask, Typer and Click, front ends with React, scraping with Playwright and Scrapy. Agentic coding with Claude Code, Codex and Cursor, backed by tests and CI.