Technical writing Research notebook
Derivations, implementation notes, and field reports from building AI systems—written close enough to the code that the inconvenient details remain visible.
4 articles
How does a global search over exponentially many label sequences collapse into a quadratic-time recurrence without changing the answer?
A derivation of exact MAP decoding from the score of a linear-chain conditional random field, with backpointers and an auditable NumPy implementation.
Read articleWhen an NER score moves, which boundary, type, context, and annotation failures actually changed—and what should be fixed next?
What evidence distinguishes a camera cut from ordinary motion, and how should per-frame boundary probabilities become usable shots?
Should a medical entity linker trust shared words, learned semantic proximity, or a deliberately hybrid candidate set?