All activity
LLMs break JSON constantly. Streaming cutoffs leave incomplete objects. Claude wraps valid JSON in markdown fences. GPT hallucinates trailing commas and unquoted keys.
Your production app crashes. You rewrite prompts 47 times.
Shim intercepts malformed LLM outputs and repairs them in sub-50ms. Real-time streaming support. Zero data persistence. Drop-in API for production AI agents.
Fixes: Truncated objects, markdown wrappers, trailing commas, unquoted keys, NaN values, schema validation.
Shim.soJSON repair API for broken LLM outputs. Sub-50ms streaming.
Andrewleft a comment
Hey everyone. Built Shim after 6 months debugging production AI agents. Same failure pattern: LLM returns invalid JSON, app crashes. **The problem:** - GPT-4 stream cutoffs (incomplete objects) - Claude markdown wrappers (```json fences) - Unquoted keys, trailing commas, NaN values - Broken production apps **The usual fixes don't work:** - Rewrite prompts 47 times → still breaks - Add 200 lines...
Shim.soJSON repair API for broken LLM outputs. Sub-50ms streaming.