✓ assert role == "Senior ML Scientist, Microsoft" ✓ assert builds == ["LLM features", "agents", "evals"] ✓ assert ships == "models into products, at Office scale" ✓ assert writing == "a GPT in pure Python. no frameworks." 4 passed in 0.00s

I get language models into products at Microsoft — designing the systems around them, fine-tuning them for the job, and wiring them into Word, PowerPoint, Excel, OneNote, and Teams. The other half of the work is knowing whether any of it is actually good, which is why I also lead evaluation for the Office Copilots: benchmarks, judge models, and the failure taxonomies that come out the other side.

Eleven plus years of shipping software/models — four as a data scientist at Penske Logistics owning freight-pricing and driver-safety systems end to end, and three before that writing C/C++ for automotive infotainment. Mostly I’ve worked on the unglamorous middle: the architecture, orchestration, and plumbing between a model that works in a notebook and one that works for millions of people.

writing

I’m building a transformer from nothing — no PyTorch or TensorFlow, just Python lists and loops — and then adding back everything real inference systems need: batching, ragged batching, a KV cache, sampling. Each post ends by proving the new version still matches the old one.

all posts →

papers

Office Comprehension Benchmark (OCB): A Benchmark for Document Comprehension across Office Documents

preprint · project lead and first author

arXiv:2607.01245 dataset code

PPT-EVAL: A Benchmark for Computer-Use Agents on PowerPoint Tasks

ICML 2026 · code released

arXiv:2606.31154 code

all papers →

elsewhere

linkedin · github · firoz@firozshaik.com