Engineering / topic
Evals
Posts from Lindy's engineering team about evals.
- Migrating from Claude to DeepSeekWe moved most of Lindy's model traffic from Claude and Gemini to DeepSeek v4 Flash and cut inference costs on those routes by about 90%. Here's the eval and rollout process that made it safe.5 min
- AI agent guardrails: how we built a validator to stop rogue actionsA separate LLM call reviews every side-effectual action before it runs. The agent has to justify it, and the validator checks that against what the user actually asked for.8 min