Data & ML Engineer in Zurich. I build data pipelines, then fine-tune and evaluate small models on top of them, and I report what doesn't work as well as what does.
10+ years in software engineering, 4+ in applied ML. MSc Information Technologies, BSc Physics.
| Project | What it is | Result |
|---|---|---|
| evidence-rag | Evidence-first multilingual RAG (EN / DE / FR): BM25, dense and hybrid retrieval, reranking, NLI citation checks, 154 tests and CI | NLI verifier cuts false citation accepts from 56% to 6%; best retrieval nDCG@10 0.805 |
| lerobot-sim-transfer-cube | ACT imitation-learning policy for the ALOHA TransferCube task, trained from scratch in simulation on an M4 MacBook with LeRobot and evaluated against the official checkpoint: same seeds, 95% intervals, two training seeds | 75.5% success vs 81.0% for the official checkpoint (200 episodes, overlapping intervals); loss curve shown not to predict success · model |
| swiss-bookkeeping-ai | LoRA fine-tune of Qwen2.5-0.5B for Swiss SME account suggestion, trained with MLX on Apple Silicon | 16.2% → 93.2% accuracy on held-out vendors, 98.6% top-3 · live demo |
| efficiency-lab | Accuracy, latency and memory of fp16, 8-bit and 4-bit variants against a larger prompted model | Fine-tuned 0.5B beats prompted 1.5B by 54 points at one third of the memory |
| bookkeeping-mcp-agent | MCP server that drafts bookings with an 8-bit model plus deterministic VAT and amount checks; never posts on its own | Stress-tested with 352 adversarial inputs, failure rates reported |
| code-model-lab | GLM-style code model trained from random weights in PyTorch, then improved with reinforcement learning | 3 → 15 of 24 unseen tasks solved after RL; seed-to-seed noise measured |
Models, datasets and demos: huggingface.co/ChadSaglam
Python, PyTorch, MLX, Hugging Face, LoRA, LeRobot, scikit-learn, RAG, SQL, PostgreSQL, ETL/ELT, Docker, Git, CI