Design · Build · Deploy · Operate

Work

Behavior Classification Inference Pipeline — 3.3× End-to-End

Bodit · 2026.02 · Led the rework

  • Python
  • NumPy
  • LightGBM
  • MLflow

Reworked the production core package that extracts 903 features from 6-axis IMU raw signal, cutting end-to-end processing time 3.3×.

3.3×end-to-end speed24h real-animal · 2.16M raw rows
433win/sec8,640 windows · was 131 · end-to-end
208 / 0pass / failfull suite · features within 1e-6

Internal LLM Data Agent Platform

Bodit · 2026.04 – present · Designed, built & operated solo

  • Claude API
  • PostgreSQL · pgvector
  • ClickHouse
  • MCP
  • Docker

An assistant that answers questions about internal data in natural language. Moved from MVP to a tool-calling agent, with the architecture decided by A/B measurement.

5 / 5compound querieslight eval, 5 cases · impossible in the old structure
730schema documentsat MVP · 95 tables · 635 columns
4surfacesMCP · Slack · REST · app API

Company Knowledge Base — Privacy-Safe Compilation Pipeline

Bodit · 2026.07 – present · Designed, built & operated solo

  • Python
  • GPU NER
  • Claude API
  • cron

A knowledge base where an LLM compiles and maintains internal conversations. Answers come from compiled documents, not from retrieval over raw text.

2de-identification layersregex + GPU NER
12ingest channelsinternal messenger
181 → 0plaintext ID numbersbefore/after rule tightening · 20 regression asserts

Threat Intelligence Data Platform — Ingest to Serving

SecuLetter · 2022.08 – 2024.07 · 3-person team · 50–80% contribution

  • FastAPI
  • AWS Lambda · ECS
  • OpenSearch
  • Svelte
  • BERT-CNN

A TI platform run end to end by a team of three: crawling, storage, search, API and UI. This is where I learned the difference between building a model and running one as a service.

5 / 6stages I builtingest → clean → store → API → UI (not search)
3team sizeinternal analysts + external customers