ToothHound

The problem

Shark-tooth hunting is a $100M-adjacent hobby trapped in tribal knowledge: Facebook groups, gut feel, and “go after a nor’easter.” Hunters drive hours on the wrong tide, misidentify a $4,000 megalodon as a common sand tiger, and — at the high end — pay seven figures for “fossil land” with no idea whether the parcel sits on the Calvert formation or a flood-zone bulkhead.

The pain is wasted weekends, blown IDs, and unbounded risk on land purchases. ToothHound makes the hidden data — tide, geology, flood zone, depth-to-bed, prior sightings — show its work.

What we ship

• Daily Hunt Score per site (0–100) fusing NOAA tides, lunar phase, wind history, USGS gauges, and active NHC storms.
• Curated hunt-site map from Florida to New Jersey with locality-specific tide-station mapping and outcome aggregates contributed by users.
• Photo-based tooth ID — a two-stage pipeline where Gemini extracts silhouette, serration, and cusplet features and a deterministic scorer ranks species with locality priors and reference exemplars.
• Field logbook and insights for logging finds, weights, and conditions, with weekly tooth-count rollups per site.
• Real-estate Buyer’s Atlas ranking 6-state coastal micro-zones by formation, expected species, dig effort, and price band.
• Parcel dossier from an address — pulls Macrostrat geology, FEMA flood zone, USFWS CBRS, Paleobiology DB occurrences, iNaturalist sightings, and USGS depth-to-bed, then runs hard-gate checks (bulkhead, off-formation, sub-economic depth) and returns a buy / no-buy verdict.
• Live land listings ranked by evidence, not list price — every listing is scored through the same geology and parcel pipeline.

What’s coming

• ID model fine-tuning loop — user feedback is already captured and a curated evaluation harness is in place; retraining and prompt-tuning is the next step. In progress.
• Push and scheduled notifications for Hunt-Score spikes — the storage layer is live, the user-facing surface is still thin. In progress.

How we built it

TypeScript on Next.js 14 (App Router) and React 18, deployed on Vercel. Data lives in SQLite (better-sqlite3) with Upstash Redis for feedback capture and rate limits. Identification runs on Google Gemini 2.5 Flash. Public data sources include NOAA Tides & Currents, NHC, USGS, Macrostrat, FEMA NFHL, USFWS CBRS, Paleobiology DB, and iNaturalist; live land listings come from RapidAPI. Maps are rendered with Leaflet.