Visit Agent

Field teams were spending 1–2 hours after every customer visit manually entering activity logs — context decayed fast, data was incomplete, and reps resented the overhead. The organisation had no reliable visit-level data to train on or forecast from. A task that should take minutes was eating hours every day.

Built a voice-first agent that lets a rep speak naturally after a visit. The system captures the audio session, extracts 6+ structured fields — customer, outcome, summary, action items, blockers, and a knowledge graph of relationships — and commits everything in under 20 minutes. No keyboard, no forms. Every session is fully traced in Langfuse with LLM-as-a-Judge evaluators for coverage and hallucination, so the team can see exactly what was captured and why.

• 01

  One-shot session tickets with a short TTL for WebSocket auth prevent replay attacks without needing a full OAuth flow — a clean pattern for any real-time backend.

• 02

  Separating live audio from post-call extraction was the right call: the audio path is fire-and-forget at low latency; the extraction jobs are async, retriable, and isolated from the voice experience.

• 03

  Langfuse prompt management in production means zero deploys to iterate on the system prompt — fetch the live version at session start and changes are immediate.

• 04

  Structured extraction with Pydantic response formats is significantly more reliable than raw prompting for typed fields — ActionItemList and summary responses parse cleanly even on messy transcripts.