Spaces:
Running
Running
metadata
title: Retro Alpha
emoji: 📺
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
tags:
- track:wood
- sponsor:nvidia
- sponsor:modal
- achievement:offgrid
- achievement:welltuned
- achievement:offbrand
- achievement:llama
- achievement:sharing
- achievement:fieldnotes
license: mit
Retro Alpha
A 90s CRT-style Indian stock market survival game (1994–2004), powered by a fine-tuned NVIDIA Nemotron-3-Nano 4B model.
Built for the 🤗 Build Small Hackathon.
🔗 Links
| 🎮 Play (main) | build-small-hackathon/retro |
| 🎮 Play (alpha) | sankalphs/retro-alpha |
| 📝 Build Blog | sankalphs.blogspot.com/2026/06/retro-alpha.html |
| 🐙 GitHub | github.com/sankalphs/retro |
| ▶️ Video Demo | youtu.be/OLJDXhos0Iw |
| 🐦 Social Post | x.com/sankalphs/status/2066602928350359694 |
🎮 How to Play
| Step | Action |
|---|---|
| Goal | Turn ₹10,00,000 into ₹20,00,000 over 10 years (120 months) |
| Review | Check the Market Watch for asset prices & trends |
| Trade | Buy/Sell any asset as a % of your portfolio using the Order Pad |
| Advance | Press Advance Month to trigger real historical events & market moves |
| Analyze | Ask the AI Advisor about your portfolio or strategy |
| Review | Get a Year-End Mentor Review for a sarcastic roast & investment lesson |
Historical Events
Asian Financial Crisis, Pokhran-II nuclear tests, Dot-com bubble, 9/11, 2004 Indian elections, and more — all influencing asset prices based on real historical data.
🏆 Badges Earned
| Badge | How |
|---|---|
| Off-Brand | Custom CRT terminal UI built from scratch (no Gradio default) |
| Well-Tuned | Fine-tuned Nemotron-3-Nano 4B on 1,500+ synthetic market scenarios |
| Nemotron | Uses fine-tuned NVIDIA Nemotron-3-Nano-4B (Q4_K_M GGUF) |
| Off the Grid | Fully self-contained Docker Space with on-device inference |
| Sharing is Caring | Infrastructure-as-code scripts open-sourced on GitHub |
| Field Notes | Detailed build log & methodology documented |
🧱 Tech Stack
Frontend → Custom CRT terminal UI (vanilla HTML/CSS/JS) served via ASGI
Backend → Python simulation engine + Gradio API
Model → Fine-tuned NVIDIA Nemotron-3-Nano 4B (Q4_K_M GGUF)
Inference → Modal GPU cloud endpoint (A10G) with deterministic fallbacks
Data → 1,500+ synthetic Indian market scenarios via zenmux API
CI/CD → GitHub Actions → HF Spaces auto-deploy
🚀 Running Locally
pip install -r requirements.txt
MOCK_LLM=1 python app.py
For LLM-powered features, set one of:
MODAL_INFERENCE_URL— Modal cloud endpointHF_API_URL+HF_TOKEN— Hugging Face Inference API
📁 Project Structure
├── app.py # Gradio app entrypoint (ASGI)
├── agents.py # LLM inference wrapper
├── engine.py # Market simulation engine
├── events.py # Historical event triggers
├── mentor.py # AI mentor review generator
├── modal_app.py # Modal GPU inference endpoint
├── download_model.py # GGUF model downloader
├── Dockerfile # HF Space container
├── requirements.txt # Runtime dependencies
├── requirements-train.txt # Training dependencies
├── config/
│ └── assets.json # Asset definitions
├── static/ # Frontend (CSS, JS, HTML)
├── schemas/ # JSON schemas for dataset validation
├── data/ # Training datasets
├── scripts/ # Dataset generation & validation
├── training/ # Modal LoRA fine-tuning scripts
└── tests/ # Test suite
🧠 Model
The game uses a LoRA fine-tune of NVIDIA Nemotron-3-Nano-4B on a custom dataset of 1,500+ Indian market scenarios covering:
- Agent decisions (730 examples) — institutional, retail, and tech-permabull personas
- News impacts (281 examples) — historical event market reactions
- Mentor reviews (255 examples) — year-end portfolio roasts with Sharpe ratios
- Guardrails (180 examples) — safety and formatting guidelines
Fine-tuned on Modal A100 40GB → exported as GGUF Q4_K_M for efficient inference.
📄 License
MIT