WhereShot

Where was this photo taken? See the answer — and how we got there.

A multi-agent forensic photo analyzer. Upload one or more photos of the same place; specialists look for tiny clues — shadows, sun angle, weather, vegetation, signage, license plates, architecture, infrastructure — and a deterministic verifier cross-checks the top guess with pvlib (sun position), Open-Meteo (weather), and Nominatim (does the address exist?).

Architecture

flowchart TD
    UI["Next.js + MapLibre<br/><i>upload · SSE viewer · map</i>"]
    API[FastAPI + LangGraph]
    UI <-->|"multipart → SSE"| API

    API --> P[ingest → preprocess → EXIF]
    P --> R{"EXIF GPS<br/>present?"}
    R -->|yes, short-circuit| V
    R -->|no| T["triage<br/><i>indoor / outdoor · low_info</i>"]
    T --> EX[extract_regions]
    EX --> SPECIALISTS

    subgraph SPECIALISTS [Specialist agents · parallel fan-out via LangGraph Send]
        direction LR
        sun[shadow_sun]:::vis
        env[environment]:::vis
        wea[weather_visible]:::vis
        sig[signage_commerce]:::tools
        lan[landmark]:::tools
        arc[architecture_culture]:::tools
        veh[vehicle]:::tools
    end

    SPECIALISTS --> H["hypothesizer<br/><i>selected · considered · rejected</i>"]
    H --> V{{"verifier · deterministic Python<br/>pvlib · Open-Meteo · Nominatim"}}
    V -->|contradiction · retry ≤ 2| H
    V -->|ok| RP[reporter → compute_confidence_status]
    RP --> OUT([Report])

    classDef vis fill:#dbeafe,stroke:#3b82f6,stroke-width:1.5px,color:#1e3a8a
    classDef tools fill:#fef3c7,stroke:#f59e0b,stroke-width:1.5px,color:#78350f

Loading

Blue specialists are pure-vision (no external lookups). Amber specialists are ReAct agents — they can call free, no-key tools to verify hunches:

• landmark — Wikipedia, DuckDuckGo, Jina Reader
• signage_commerce — Wikipedia, DuckDuckGo, country_info
• architecture_culture — Wikipedia, DuckDuckGo, country_info, plate_format
• vehicle — plate_format, country_info, DuckDuckGo

A step budget (≈5 tool-calling rounds, hard kill at 20 graph steps) prevents runaway loops. A budget-exhausted specialist emits honest "couldn't determine" evidence at confidence 0 instead of crashing the pipeline.

Backend — Python 3.11+, FastAPI, LangGraph + LangChain, OpenRouter (one API key for all LLMs). Per-role model routing: Gemini Flash for coarse buckets, Claude Haiku for lighter ReAct, Claude Sonnet for richer reasoning, Gemini Pro for landmark geography.

Frontend — Next.js 16 (App Router), Tailwind 4, MapLibre GL JS + OpenFreeMap vector tiles, Zustand.

Quickstart

cp .env.example .env
# add OPENROUTER_API_KEY (the only required key)
just install
just dev
# open http://localhost:3000

Or with Docker:

cp .env.example .env
just up

CLI without the web UI:

just analyze path/to/photo.jpg

Responsible use

Estimates are probabilistic and can be wrong. Intended for journalism, education, and research. Not for tracking private individuals.

License

MIT — see LICENSE.