Track 2, gxinxing, AMD ROCm Local Quantitative Investment Assistant

[gxinxing]

AMD ROCm Local Quantitative Investment Assistant

A fully local quantitative investment agent: natural language → strategy DSL → deterministic validation → backtest → independent risk-agent veto, the whole chain running on an AMD Radeon GPU (ROCm 7.2.1) — no dependency on any cloud model.

• 🧠 Qwen2.5-7B + LoRA fine-tuning, vLLM local inference (32K context)
• ✅ 285 tests passed, 24/24 domestic-market evals, 98.1% token accuracy
• 🔒 Data & inference stay local; the risk agent holds an independent veto
• 🎮 Three verification paths:
  1. Online Open WebUI — register & chat (link below)
  2. One-command local reproduction — python demos/run_track2_demo.py (no GPU required)
  3. 4m23s real-run demo video — output/video/track2_demo_1080p_ava.mp4 (English narration, real AMD GPU)

Links

• Repository: https://github.com/gxinxing/Radeon-hackathon-2026-07
• Live showcase (landing): https://autoquant-landing.vercel.app
• Open WebUI (register & chat): https://bones-fans-imposed-rio.trycloudflare.com
• Model weights (ModelScope): https://modelscope.cn/models/gxinxing/qwen-trader-cn-merged

Official deliverables alignment

#	Requirement	Where
1	Project Specification Document (PDF)	output/pdf/AMD_Quant_Assistant_Project_Specification.pdf
2	Source code + README + env/startup guide	entire repo; see README ## One-command reproducible demo
3	Demo video (real run on AMD GPU)	output/video/track2_demo_1080p_ava.mp4 (4 min 23 s)
4	Supplementary material (PPT / Poster)	output/pdf/track2_submission_ppt.pdf + output/poster/

Reproduction (on an AMD ROCm machine)

Full step-by-step guide in README.md. Quick start: follow train_qlora.py → merge → serve or run the CPU-only demo python demos/run_track2_demo.py to see the full DSL → backtest → risk-veto pipeline without a GPU.