Results
ArchiveAgent
I fine-tuned a small model to read runway photos. It beat much larger models.
| ArchiveAgent | 56.4% | |
| Kimi K3 | 44.9% | |
| Claude Opus 5.5 | 28.6% | |
| Before training | 12.2% |
How often each model named the house, season and year correctly. 287 new runway photos, October 2026. See the results.
Work
Research
- Genesis 03: where 380 brand links from ChatGPT and Gemini shopping answers actually went.
- Reverse Engineering How Muse Shops: how Meta's AI agent runs product catalog search. Evidence on GitHub.
- Shopping Answers: 459 shopping questions and 457 captured AI answers across 31 categories.
- Local image models on fashion photos: four open image models tested on 86,874 runway photos. The best found the right brand 74.4% of the time on one home GPU.
Projects
- ArchiveAgent: a personal project. Theia, a fine-tuned Kimi K2.6, reads a fashion photo and names the house, season and year. Benchmark on Hugging Face.
Earlier
Heat2Power, Data Analysis Intern (Ann Arbor SPARK Digital Summer Clinic), summer 2025: public energy data analysis at a thermophotovoltaic (TPV) semiconductor startup.
Pinecone Pioneers, 2025 cohort.
Contact
LinkedIn · GitHub · Hugging Face · X