Picnix

buddy trip planner with a plan that works!

Comment

Picnix: Your Verified Weekend Trip Planner

nvm the rick roll, i couldn't get the video done :) but the project is cool, I promise - use it at picnix.aswinpradeepc.com

Inspiration

I'm a dev working away from home in Kochi. Every Friday, the same ritual: someone in the flat asks "so… anywhere this weekend?" and we all open forty browser tabs. Half the places are closed Sundays. A third are too far for a day trip. The "hidden gem" from a 2019 blog post is permanently shut. By the time anyone has a plan, it's Saturday afternoon and we're ordering biryani in.

I tried asking a chatbot once. It confidently invented a waterfall and told me a 3-hour drive took 45 minutes. That was the moment: I didn't need an AI that talks about trips. I needed one that builds them — and proves its work.

What it does

Picnix turns one short conversation into a verified, mapped, minute-by-minute day trip. It extracts your constraints (max 3 questions), computes the area you can actually reach, pulls live Google Places candidates, and validates every single one against real opening hours, closure flags, and actual Routes API travel times — you're never shown a place the agent can't prove you can visit.

You pick 1–3 stops from a validated gallery (human-in-the-loop, always), and it builds a multi-waypoint round trip with real ETAs, reasoned dwell times, and food planned along your actual route geometry. Want a change? Say "swap the beach for the museum and leave at 9" — it re-plans through the same validation gauntlet. One tap exports to Google Maps for turn-by-turn.

And because agents fail in interesting ways, Picnix ships a second agent: the Trip Auditor, built on the Arize Phoenix MCP server, that answers "why did my plan drop a destination?" by reading the planner's own traces.

How I built it

  • Core Architecture: An 8-node LangGraph state machine powered by Gemini 3.1 Pro (reasoning: intent, routing, validation, editing) and Gemini 2.5 Flash (prose) on Vertex AI.
  • Ground Truth Layer: Google Maps Platform (Places, Routes, Geocoding).
  • Observability: Every node, LLM call, and tool invocation streams to a self-hosted Arize Phoenix collector via OpenInference auto-instrumentation.
  • Storage: PostgreSQL backs both user accounts and LangGraph checkpoints, so interrupted plans survive restarts.
  • Deployment: The whole stack — app, Phoenix, Postgres — deploys to a GCP Compute Engine VM with one docker compose up.
  • Engineering Quality: 175 tests, 13 ADRs.

Challenges I ran into

  • Hallucination doesn't die quietly: My first composer invented place names. The fix became the architecture: validate the structured plan before prose exists (Python checks + a Gemini semantic pass), then force the composer to emit a claim audit — every sentence traced to a verified data field, unverified claims stripped. I even swapped two graph nodes mid-build (ADR-006) to make validation come first.
  • Privacy for a trace-reading agent: Phoenix API keys grant org-wide access, and traces contain user locations and chats. Prompt-level "please only look at your own data" isn't security. So I made the tool surface the boundary: regular users get exactly two database-validated tools scoped server-side to their own trips via session.id span attributes. Prompt injection can't widen a toolset that was never there.
  • Production papercuts: Gemini 3.1 Pro is global-endpoint-only; JSON mode without a response schema lets keys drift; quota spikes needed centralized retry/backoff. Each one is now an ADR.

Accomplishments that I'm proud of

  • An agent that refuses to lie — the claim-audit pipeline means hallucinations structurally can't reach the user.
  • An agent that audits the agent — real MCP integration, not a dashboard screenshot.
  • Per-user trace scoping that treats security as architecture.
  • A genuinely shippable product: auth, email verification, durable checkpoints, one-command deploy — built solo, with every architectural decision documented.

What I learned

  • Observability isn't an afterthought: Phoenix traces found my bugs before users did, and then became a feature.
  • Human-in-the-loop is a design principle, not a checkbox: The graph never silently swaps a destination.
  • The best guardrail against hallucination isn't a better prompt: It's a pipeline where unverified facts have nowhere to go.

What's next for Picnix

  • Multi-day trip planning (the graph already routes around it).
  • User-driven stop removal and edit-time place additions.
  • Arize AX for production monitoring at scale.
  • Traffic-aware re-planning.
  • WhatsApp itinerary sharing — because the group chat is where every Kochi weekend actually starts.

Built With

Share this project:

Updates