Currently building AX-Office.ai, an on-prem LLM platform powering five production apps inside a security company.
|
Local-first LLM apps, RAG, OCR, ASR, agent workflows, FastAPI services. |
Eval harnesses, model regression checks, observability, deployment discipline. |
PyTorch, SFT, distributed training, benchmarking open-source models. |
| Project | What it is | Stack |
|---|---|---|
| AX-Office.ai | On-prem AI platform with five production apps and zero external LLM calls. | vLLM RAG ASR OCR FastAPI |
| GPT-2 124M from scratch | Pretrained a 124M GPT-2 on 18B curated tokens, then instruction-tuned it. | PyTorch pretraining SFT |
| Ajax shot analyzer | Computer-vision pipeline for football shooting technique and 3D skeleton visualization. | MediaPipe OpenCV Three.js |
| ESRGAN with uncertainty | Master's thesis on medical-image super-resolution with calibrated uncertainty. | PyTorch ESRGAN Bayesian DL |
| Area | Tools |
|---|---|
| LLM systems | Python vLLM Hugging Face Anthropic SDK RAG |
| Training | PyTorch SFT distributed training model benchmarking |
| Backend | FastAPI Docker Linux GitHub Actions |
| Data | PostgreSQL pgvector evaluation datasets |
| Interests | agent reliability local AI football analytics |