Hoàng
Building biomedical NLP systems
N–07

Current research log

Now is a small set of questions under active pressure.

This is a snapshot, not a list of achievements. It records what is taking most of my attention, the latest concrete change to each system, and the question that should shape the next experiment.

Cadence
Updated when the work changes
Current cycle
July 2026
01

Three workstreams

Closest to the keyboard

  1. 01

    Biomedical NLP competition pipeline

    case study
    Current stage
    Pipeline decomposition and evaluation design
    Most recent milestone
    Separated span recognition, assertion handling, candidate retrieval, and reranking into independently inspectable stages.
    Next technical question
    How much linking error comes from candidate recall, and how much remains after a correct concept enters the top-k set?
  2. 02

    POV video understanding

    case study
    Current stage
    Boundary-aware preprocessing
    Most recent milestone
    Defined a timestamp-preserving path from FFmpeg decoding through TransNetV2 boundaries to SigLIP2 keyframe deduplication.
    Next technical question
    Can local similarity filtering remove redundant frames without discarding brief actions near uncertain shot boundaries?
  3. 03

    2026 Student AI Olympiad preparation

    case study
    Current stage
    Mathematics and reference implementations
    Most recent milestone
    Established a repeatable derive–implement–interrogate–verify format for each algorithm module.
    Next technical question
    Which small examples best expose the difference between max-product decoding and sum-product inference?
02

Study queue

Foundations in rotation

  1. 01

    Structured prediction

    Viterbi, forward algorithm, CRF gradients

  2. 02

    Retrieval systems

    BM25, bi-encoders, score fusion, reranking

  3. 03

    Numerical foundations

    log-sum-exp, stable softmax, gradient checks

  4. 04

    Video representation

    shot transitions, temporal sampling, frame diversity

03 / NEXT REVISION

What would make this page change?

A retrieval recall study, a reviewed POV boundary slice, or a verified algorithm module would each replace an intention above with evidence. Until then, the open questions stay visible.