Building AI systems that run in production, not notebooks.
- Fine-tuning — adapting open models to real workloads, LoRA / parameter-efficient methods, and evaluating whether the result actually improved
- Inference & serving — routing between models, context handling, compression, quantization, and the cost/latency tradeoffs that decide what ships
- Evaluation — benchmark harnesses and regression suites, because "it feels better" is not a result
- Knowledge graphs — modelling entities and relationships, record linkage and entity resolution, querying graphs to answer causal questions
- Agents & tooling — tool-calling agents, MCP servers, and CLIs that developers actually keep installed
- Data & telemetry pipelines — collection, transformation, and the analytics layer on top
- Infrastructure — Docker, Linux, cloud deployment, and keeping all of the above running
- Shipping — real users, real constraints, real failure modes
- Computer Vision — object detection and segmentation
- Deep Learning — CNNs and transformers, training and evaluating networks end to end
- Classical ML — regression, classification, clustering, feature engineering
ML/DL PyTorch Transformers Hugging Face scikit-learn OpenCV NumPy pandas
Systems Python Go TypeScript FastAPI Docker Linux Git React