01
Multimodality
Vision-language systems that judge content, not just caption it.
- Visual-Qwen↗ · CLIP + Q-Former + Qwen3 4B, 92% eval accuracy, trained on an H200
- MicroMARC · vision-language model that flags cognitively degrading short-form video
Test what others assume.
Independent research across intelligence, compute, and data. We run the tests ourselves and publish the code, methods, and raw numbers.
$376K
Google TPU Research Cloud compute grant
4
accelerator platforms benchmarked
2,000+
labeled video clips published
66
public repositories
On-device inference, measured
Methodology and raw logs: snapdragon-vs-m5, snapdragon-vs-mediatek
Apple M55,604
Snapdragon X2 Elite1,921
Apple M5144
Snapdragon X2 Elite75
M5 ANE, int815,593
X2 HTP, fp1615,205
Snapdragon 8 Elite4,307
Dimensity 9500s1,470
Research
02
Accelerated computing
One PyTorch workload, profiled across GPUs, TPUs, NPUs, and phones.
- De-mystifying PyTorch for ASICs · accepted talk, PyTorch Conference Europe 2026
- Edge silicon benchmarks↗ · Apple M5, Snapdragon X2 Elite, Dimensity 9500s, head to head
- will-it-asic↗ · model compatibility across TPU, Trainium, Inferentia, and Gaudi
03
Data engineering
Datasets and pipelines built to be rerun, not just cited.
- Multimodal Sludge Dataset · 2,000+ labeled video clips, published on Kaggle
- ts-jobspy↗ · job market scraper for LinkedIn, Indeed, Glassdoor, and ZipRecruiter
- chorus-searxng↗ · self-hosted private search with AI answer synthesis
Selected work
All 66 repositories ↗- vqwen-qformerVision-language model: CLIP + Q-Former + Qwen3 4B, trained on an H200.
- snapdragon-vs-m5Reproducible head-to-head: Apple M5 against Snapdragon X2 Elite.
- snapdragon-vs-mediatekOn-device benchmark: Snapdragon 8 Elite against Dimensity 9500s.
- will-it-asicChecks model compatibility across TPU, Trainium, Inferentia, and Gaudi.
- poco-phone-ai-trainingTraining and inference on a retail phone NPU over SSH.
- local-agentic-rag-with-qwen3Local agentic RAG on Qwen3 embeddings, reranker, and LLM.
Independent.
No vendor allegiance. We buy or rent the hardware we test.
Open.
Methods, code, and raw logs ship with every result.
Measured.
Claims come from benchmarks we ran, not spec sheets.