DEV Community

Igor Ganapolsky
Igor Ganapolsky

Posted on • Originally published at igorganapolsky.github.io

Building an AI Trading System for the Tetrate Buildathon

#tetrate #buildathon #tars #aitrading

The Buildathon

The Tetrate AI Buildathon challenges participants to build or enhance applications using TARS (Tetrate Agent Router Service) — an AI gateway that routes LLM calls across multiple providers.

We're bringing an existing system: an autonomous AI trading system that executes SPY iron condor options strategies with $100K in paper capital. The system already had multi-model routing built locally. TARS lets us move that routing to a centralized gateway with features we can't replicate locally.

What We Already Had

Before the buildathon, our system included:

  • Budget-Aware Model Selection (BATS) — routes tasks to the cheapest capable model ($25/month vs $500+)
  • 5 LLM models across 3 providers (DeepSeek, Mistral, Kimi K2, DeepSeek-R1, Claude Opus)
  • Feedback-driven context pipeline — Thompson Sampling + LanceDB + MemAlign for continuous learning
  • 84 GitHub Actions workflows — self-healing CI that monitors, fixes, and learns autonomously
  • 170+ documented lessons — every failure recorded, indexed, and searchable via semantic search

What TARS Adds

Feature Before (Local) After (TARS)
Fallback routing Code-level fallback chain in model_selector.py Gateway-level auto-failover across providers
Budget enforcement Local tracking, resets on restart Server-side per-token budgets, persistent
Traffic splitting Not possible A/B test model quality (e.g., 90% Kimi K2 / 10% new model)
Telemetry Manual logging Centralized request logs, usage dashboards, cost tracking
MCP profiles N/A Curated tool subsets for different trading agents
Key management Multiple env vars per provider Single TARS key, BYOK for each provider behind the gateway

The integration point is minimal — two environment variables:

LLM_GATEWAY_BASE_URL=https://api.router.tetrate.ai/v1
TETRATE_API_KEY=sk-your-key
Enter fullscreen mode Exit fullscreen mode

Every OpenAI-compatible call in the system routes through TARS with zero code changes.

Architecture

LLM Gateway Architecture

Trading Pipeline

Feedback Pipeline

What We Built Today

In one buildathon day:

  1. README rewrite — documented the real architecture with TARS integration for judges
  2. 3 PaperBanana diagrams — auto-generated publication-quality architecture visuals via Gemini
  3. 2026 SOTA comparison — researched how our feedback pipeline compares to Mem0, OpenAI Agents SDK, and state-of-the-art agent memory systems
  4. 4 blog posts — this one, plus deep-dives on feedback pipelines, TARS routing, and PaperBanana automation
  5. 3 Claude Code skills — /generate-diagram, /generate-plot, /update-diagrams for repeatable diagram generation

Key Insight

The biggest value of TARS isn't replacing what we already built — it's centralizing it. Our local model_selector.py does budget-aware routing well. But TARS adds the layer above: gateway-level failover, server-side budget persistence, traffic splitting for model evaluation, and a telemetry dashboard that works across all our agents without custom instrumentation.

For a trading system where reliability directly equals money, that centralization matters.

Built for the Tetrate AI Buildathon. Full source code at github.com/IgorGanapolsky/trading.

Top comments (0)