← All projectsOngoing
Personal R&D
Function-Calling Support Assistant
A support-ticket agent that decides when to call real Python functions — order lookup, refund calculation — instead of guessing.
AI / ML
Overview
Built the core agent loop from scratch: send messages to the LLM, let it emit structured tool calls, execute the matching Python function, feed the result back, repeat — the actual mechanics behind 'agentic' behavior.
What I did
- •Implemented an agent loop with tool-call execution and a bounded turn limit to avoid runaway loops
- •Defined JSON-schema tool contracts (lookup_order, calculate_refund) the model calls with structured arguments
- •Applied a real refund policy (14/30-day tiers) as the deterministic logic behind the AI-facing tool