← Back to work
Applied LLM & RAGLive

Vehicle DIY Guide

An agentic RAG system that builds vehicle-specific repair guides from manuals, forums and video — with a live AI mechanic you can ask follow-ups.

View codeLive demo
Vehicle DIY Guide — product screenshot
WrenchAI — enter your exact vehicle + problem to generate a step-by-step guide.
5
source types fused per guide
$0
cost on cached repeats
2
cost-tiered Claude models
FastAPINext.jsClaudepgvectorRedisPythonTypeScript

The problem

Generic AI answers are useless for car repair — the right steps depend on the exact year, make, model and engine. Owners need guidance grounded in real, vehicle-specific sources, not a plausible-sounding hallucination.

The approach

A knowledge-builder agent fans out across NHTSA technical service bulletins, Reddit (PRAW), YouTube transcripts and web search in parallel, then a synthesizer merges them into a single vehicle-specific, step-by-step guide with safety tiers. Guides are embedded and cached in Postgres+pgvector so repeat queries are instant and free. An intent-routed chat session lets the user ask real-time follow-ups grounded in the retrieved context.

Architecture

Vehicle input (year/make/model/engine + problem)
        │
        ▼
Knowledge-builder agent ── parallel ──▶ NHTSA TSB · Reddit · YouTube · Web search · Images
        │
        ▼
Synthesizer (Claude Sonnet) ─▶ safety-tiered, step-by-step guide
        │
        ├─▶ embed + cache ▶ Postgres + pgvector   (repeat queries = instant, $0)
        │
        ▼
Guide session (intent routing) ─▶ AI mechanic chat (Claude Haiku) grounded in retrieved context

Key decisions & trade-offs

  • ▸Claude Sonnet for one-shot guide synthesis (quality), Claude Haiku for chat (cheap, fast) — cost-tiered by task.
  • ▸pgvector over a hosted vector DB: one Postgres does storage + retrieval + caching, less infra to run.
  • ▸Aggressive caching — the second person asking about the same repair pays nothing and waits milliseconds.
  • ▸Safety tiers classify repairs (cosmetic → mechanical) so the UI can warn before risky jobs.

Results / impact

  • ✓Produces a vehicle-specific, safety-tiered guide in a single pass from 5 parallel sources.
  • ✓Caching drops repeat lookups to instant, $0 responses.
  • ✓Cost-tiered LLMs keep the price of each generated guide low without sacrificing quality.

Roadmap

  • ▸Add vehicle-recall alerts sourced from the NHTSA feed.
  • ▸Explore a scale-to-zero serverless backend (Bedrock) to cut idle hosting cost.
← All projects