tusharjayanti

Vox Agent — AI skill for Claude Code

AI community

LLM agent for customer support with inline evaluation, hallucination detection, retry/fallback logic, and RAG grounding.

How to install Vox Agent

This entry records only its repository, not the path inside it, so there is no exact command to give. Open tusharjayanti/vox-agent and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Vox Agent does

LLM agent for customer support with inline evaluation, hallucination detection, retry/fallback logic, and RAG grounding. Built with Claude (Sonnet + Haiku), FastAPI, asyncpg, and pgvector.

Alternatives in AI

Browser Search — A skill for AI agents: search the web with SearXNG, browse with Camofox, bypass protections with CloakBrowser 499 ★
Vox AI Motion Graphics Generator — 🎬 Turn any topic into a finished Vox-style paper-collage explainer / motion graphics video — script, collage 202 ★
Gh PR Review — GitHub CLI extension that adds full inline PR review comment support — view, navigate, reply to, and resolve r 174 ★

README

vox-agent

A customer-support AI agent with **inline response evaluation**, hallucination detection, and automatic retry/fallback. Built with Claude Sonnet + Claude Haiku (as judge), FastAPI, and Postgres.

What it does

vox-agent is a FastAPI chat service where every agent response is evaluated inline by a second model before reaching the user. Claude Sonnet generates the response; Claude Haiku scores it against a rubric covering relevance, groundedness, and hallucination risk. Responses that fail get one targeted retry with a failure-specific correction; unrecoverable failures return a safe fallback message with a lookup ID. Every turn and every evaluation is persisted to Postgres for historical analysis.

Why it exists

Most LLM demos skip the hardest part of running agents in production: **knowing when the agent is wrong**. Post-hoc eval runs on sampled traffic miss live failures; system-prompt engineering only goes so far. vox-agent treats evaluation as a first-class request-time concern, with the cost shape (Haiku as judge) that makes evaluating every turn affordable in real-world deployments.

Architecture

┌──────────────┐
│ POST /chat   │
└──────┬───────┘
       ▼
┌────────────────────────────────────────┐
│ agent.chat()                           │
│  1. Load history (memory.py)           │
│  2. Generate (Claude Sonnet)           │
│  3. Evaluate:                          │
│     ├─ Heuristics (sync, free)         │
│     └─ LLM judge (Claude Haiku, JSON)  │
│  4. If retry: targeted correction → re-gen + re-eval
│  5. If escalate: hardcoded fallback    │
│  6. Persist turn + evaluation          │
└──────┬─────────────────────────────────┘
       ▼
┌──────────────┐
│   Postgres   │  turns + evaluations (asyncpg)
└──────────────┘

**Two-layer evaluator.** Heuristics (length, refusal detection, hedge density, off-topic vocabulary) run first — deterministic, free, short-circuit the judge when a response is obviously bad. The LLM jud