Computer Science · AI · mathematical practice
I study intelligent systems by making their intermediate steps visible.
I’m Hoàng, a Computer Science student focused on artificial intelligence, natural language processing, machine learning, and the mathematics underneath them.
This site is part portfolio and part working notebook. It holds systems I am building, explanations I needed in order to understand them, and experiments whose conclusions are still allowed to change.
Current orientation
What I am trying to understand
- 01
Language as structured data
Biomedical NLP is the current center of gravity: sequence labelling, assertion context, retrieval, and linking mentions to controlled vocabularies.
- 02
Models as mathematical objects
I want to understand an objective well enough to derive its recurrence, identify its assumptions, and explain its failure modes without hiding behind an API.
- 03
Experiments as evidence
Small baselines, fixed evaluation slices, and error ledgers come before impressive demos. An unreported result is better than an untraceable one.
Understanding should survive the move from paper to code.
My preferred unit of work is a small, auditable artifact: a derivation, a reference implementation, a visual trace of the internal state, and tests that can be checked by hand.
Operating principles
How I work
- 01Derive before optimizing
- Write down the state, objective, and complexity before reaching for a faster implementation.
- 02Keep interfaces inspectable
- Separate recognition, retrieval, and ranking so a failure can be assigned to the stage that produced it.
- 03Prefer tiny decisive tests
- Use hand-computable examples, invariants, and adversarial cases to expose implementation mistakes early.
- 04Publish the unfinished edge
- Record failed approaches and next questions alongside what currently works.
The useful parts are still moving.
The project pages document architecture and negative evidence. The writing archive develops the mathematics and implementation details. The Now page records the questions currently closest to the keyboard.