TLDR Tech / Crypto
Lessons from Building AI Agents That Handle Real Money
Ask Gina's AI agent platform has processed $5M+ in volume across 100K+ transactions on Polymarket, Hyperliquid perps, and 12+ chains since early 2024. The team's primary finding is that LLMs cannot reliably handle precise financial math such as basis-point position sizing under leverage, so they shifted to a model that writes deterministic code rather than computing outputs directly, combined with a modular pipeline separating scanner, filter, executor, and monitor steps. Filesystem-based memory outperformed vector databases and RAG pipelines in practice. The single most effective error-prevention guardrail was requiring a plain-English strategy plan before any code generation.
Read the full story