MISTRAL AI WORLDWIDE HACKATHON — BUILDING QUANTBRIEF IN 48 HOURS

The Hackathon

On February 28, 2026 I participated in the Mistral AI Worldwide Hackathon 2026 (Anything Goes track). In 48 hours, I built QuantBrief — a full-stack multi-agent market intelligence platform that acts as a personal Chief Investment Officer for retail investors.

The idea came from a simple observation: retail investors face an impossible information asymmetry. Over 300 financial articles are published every hour, SEC filings run 80–120 pages, and a Bloomberg Terminal costs $25,200/year. QuantBrief bridges that gap with AI-powered analysis on a schedule you define.

Why Multi-Agent

A single LLM call can't handle financial intelligence properly. News screening requires speed (thousands of articles). Filing analysis requires depth (full document context). Reasoning requires chain-of-thought. Synthesis requires cross-source correlation. Each task has fundamentally different requirements.

The solution: a 5-agent orchestrated pipeline where each agent uses a Mistral model optimized for its specific role:

① News Screener     → Ministral 3B        (~50ms/article, batch JSON)
② Filing Analyst    → Mistral Large 3      (256K context, full 10-K)
③ Reasoning Engine  → Magistral Medium     (CoT impact assessment)
④ Synthesizer       → Mistral Large 3      (cross-source brief)
⑤ Voice Agent       → ElevenLabs TTS       (5 languages)

The 256K Advantage

The most impactful technical decision was using Mistral Large 3's 256K context window for SEC filing analysis. Traditional approaches chunk documents into 4–8K segments and use RAG for retrieval — this introduces information loss and breaks cross-referencing between sections.

With 256K tokens, the entire 10-K filing fits in a single call. The model sees the full document: risk factors referencing revenue figures referencing forward guidance. No chunking artifacts, no retrieval misses, no lost context.

This is particularly powerful for 8-K filings (material events) where a single sentence in a 3-page document can move a stock 10%.

Architecture Decisions

Model selection per agent — Not every task needs the largest model. News screening with Ministral 3B runs at ~50ms per article (30 articles batch), while filing analysis with Mistral Large 3 takes 15–30 seconds but delivers institutional-quality analysis. This asymmetry is intentional: speed where it matters, depth where it counts.

Scheduled automation — Most AI finance tools are on-demand. QuantBrief generates briefs on a schedule (every 4h, daily, weekly) via APScheduler with PostgreSQL persistence. This transforms it from a tool into an autonomous analyst.

Free data sources only — SEC EDGAR, yfinance, RSS feeds, FRED. No paid APIs. The entire data pipeline costs $0 to operate — the only cost is Mistral API tokens.

What I Built in 48 Hours

• Backend — FastAPI with 25+ endpoints, JWT auth, PostgreSQL, Redis, rate limiting, security headers
• Frontend — React 19 + TypeScript dashboard with real-time WebSocket pipeline progress, interactive charts (Recharts + Lightweight Charts), multilingual UI (5 languages via react-i18next)
• Pipeline — 5-agent orchestrator with stage-level error handling and W&B experiment tracking
• Infrastructure — Docker Compose, GitHub Actions CI/CD, deployed to Vercel (frontend) + Railway (backend)
• Tests — 33 automated tests covering auth, portfolio, watchlist, and schedule isolation

13,500+ lines of code across 95+ files. Full-stack, production-ready, deployed and live.

Takeaways

• Multi-agent architectures are worth the complexity when tasks have genuinely different compute profiles — don't force one model to do everything
• 256K context windows change the game for document analysis — RAG is a compromise, not a solution
• Hackathons are the best forcing function for shipping fast — constraints drive better decisions
• The Mistral model ecosystem is mature enough to build real products on — Ministral 3B for throughput, Large 3 for depth, Magistral for reasoning