GamePulse — AI-Powered Roblox Game Analytics & Luau Code Generation

Built for the Mistral AI Hackathon | Powered by mistral-small-latest | Traced with W&B Weave

GamePulse is a unified platform that combines AI-powered game analytics with an intelligent Luau code generator — both powered by Mistral AI with self-improving prompt pipelines and full observability through Weights & Biases Weave.

Live Demo: https://web-production-fe47d.up.railway.app

---

What It Does

1. GamePulse Dashboard — AI Game Analytics

A consulting-grade analytics dashboard for Roblox developers. Enter any Roblox game and get instant AI-powered analysis across 10 dimensions:

Feature	What Mistral Analyzes
Portfolio Overview	AI-contextualized stats with health score
Daily Briefing	Personalized morning briefing with quick wins
Deep Dive Insights	Game health diagnosis, growth actions, competitor notes
Game Loop Analysis	Core loop rating, retention hooks, bottlenecks, monetization gaps
Competitor Radar	Competitive positioning, threats, feature gaps, "steal this" recommendations
Trend Analysis	Platform trends mapped to your specific games
Revenue Estimator	Monetization analysis with revenue percentile ranking
Social & Marketing	AI-generated tweets, Discord posts, YouTube descriptions
Update Roadmap	4-week prioritized development plan
AI Chat	Free-form conversation about your games with full context

Every AI call is traced through W&B Weave with @weave.op() decorators for full observability.

2. Luau Copilot — Self-Improving Code Generator

An AI code generation system for Roblox Luau that improves itself through an automated evaluation pipeline:

• Fine-tuned models: SFT (Supervised Fine-Tuned) and RFT (Reinforcement Fine-Tuned) LoRA adapters on Devstral
• 5 automated scorers: Syntax validation, Roblox API correctness, bug detection, code quality, and LLM-as-judge task completion
• Self-correction loop: Generate → Score → Fix → Re-score (up to 3 rounds)
• Agentic pipeline: Full generate → score → self-correct → deploy-to-Studio flow
• Roblox Studio plugin: Live code injection from web UI directly into Studio

---

The Self-Improvement Pipeline

The core innovation — an automated loop where Mistral improves its own system prompts:

FOR each iteration:
  1. GENERATE  → Mistral generates Luau code for 15 test tasks
  2. EVALUATE  → 5 scorers grade each output (traced in Weave)
  3. ANALYZE   → Identify dimensions below threshold
  4. IMPROVE   → Mistral rewrites its own system prompt targeting weaknesses
  5. SAVE      → Version the prompt (v1→v7) + log scores
  6. REPEAT    → Run again with the improved prompt

Results: 7 prompt versions evolved, with scores tracked in results/iteration_log.json. The champion prompt (v7) feeds back into both the code generator and the analytics dashboard.

Champion Prompt Evolution

The analytics prompts were evolved through a genetic algorithm:

• Genome 83e98d52 — 83.9% fitness (97.9% accuracy, 97.5% completeness, 80% insight quality)
• 6 genes: identity, analysis framework, format rules, few-shot examples, interaction patterns, benchmark context
• Survived 3 generations of crossover + AI-driven mutation

---

Architecture

gamepulse/
├── webapp/                    # Django application
│   ├── core/                  # Game analytics (10 AI features)
│   ├── copilot/               # Luau code generator + Studio integration
│   └── templates/             # Tailwind CSS dark-mode UI
├── scorers/                   # 5 automated code scorers
├── pipeline/                  # Self-improvement loop (eval→analyze→improve)
├── training/                  # SFT + GRPO/RFT training scripts
├── src/                       # Generator core + champion prompt
├── configs/prompt_versions/   # v1.txt → v7.txt (evolved prompts)
├── results/                   # Iteration logs + forge corrections
├── studio_plugin/             # Roblox Studio Lua plugin
└── scripts/                   # Pipeline runners + W&B report generation

Stack: Django · Mistral AI (mistral-small-latest) · W&B Weave · Tailwind CSS · Railway

---

Mistral AI Usage

Every AI feature runs through Mistral's chat API:

• mistral-small-latest for all analytics, code generation, self-correction, and prompt evolution
• Champion prompt system — evolved prompts with diagnostic frameworks, benchmark context, and few-shot examples
• LLM-as-judge — Mistral evaluates its own code output for task completion scoring
• Meta-prompting — Mistral rewrites its own system prompts based on evaluation feedback

---

W&B Weave Observability

All Mistral calls are traced with @weave.op():

• Every API call logged with inputs, outputs, latency, and token usage
• Evaluations run through weave.Evaluation with 5 scorer dimensions
• Model versions auto-tracked via weave.Model (prompt changes = new version)
• Dashboard: wandb.ai/carpediemhari-n-a/gamepulse

---

Quick Start

git clone https://github.com/Hcoder10/gamepulse.git
cd gamepulse
pip install -r requirements.txt

# Set environment variables
cp .env.example .env
# Edit .env with your MISTRAL_API_KEY and WANDB_API_KEY

# Run the web app
cd webapp
python manage.py runserver

# Run the self-improvement pipeline (optional)
python scripts/run_loop.py -n 5

---

Team

Built by Hari for the Mistral AI Hackathon.

---

All AI features powered by Mistral AI · Observability by Weights & Biases Weave