Case 01PASSION PROJECT · LIVE AT RIVERLINE.APP

Predicting the hatch before it happens

Predictive AI that scores fishing conditions by comparing live USGS flows, water temp, snowpack and weather against years of historical hatch and fly shop data, for 16 rivers across Montana, Idaho, Wyoming and Canada.

RoleSolo passion project

StackClaude Code, React, Cloudflare, Supabase, Groq AI

StatusLive

Read4 min

1 / 5 · Riverline on desktop and phone

Step 01 / 05
Use case

The user moment

It's Tuesday night and you're going on a 3-day trip to a river you've never fished on Saturday. You check the fly shop report, the USGS gauge, water temp, the weather, and a forum post from 2019. None agree, all the information is fragmented and relative, and none can tell you what to really expect.

Step 02 / 05
Problem

The problem

The information exists, but it's scattered across 8 fly shop reports, USGS streamflow, this year's snowpack, historic reports, hatch timing, water temp and old forums. And all of it is relative. You feel it most when scouting new water you're excited to fish, or planning months ahead, when today's report doesn't help.

Step 03 / 05
Validated

How I validated it

I'm the userI fish the Bitterroot and the Missouri, and talked the idea through with 10+ close friends who fly fish before I built it. I used it myself before trusting anyone else to.

Field test5+ rivers this summer. On the Elk River in Fernie it called the elevated flows and the hatch.

Expert checkA fly fishing guide and a lifelong angler said the picks matched what they'd recommend.

Step 04 / 05
Decided

Decisions I made

What I left out matters as much as what I built.

ConsideredMaps and float planning

ChoseLeave them out

onX and TroutRoutes do maps and float planning well. My edge is the forecast.

ConsideredToday's conditions only

ChoseForecasts 3+ months out

Trips get planned weeks ahead.

ConsideredEvery river everywhere

ChoseThe rivers I know best

I can verify every answer myself.

ConsideredTrust the AI's fly data

Chose3 AI judges, 3 retries, then me

Bad recommendations, wrong fly names and inaccuracies kill trust fast. 98% of 500+ flies came out correct.

Under the hood

How it's built

Rules do the scoring. AI explains it.

CollectLive feeds

USGS and Environment Canada flows, snowpack and weather, refreshed on a schedule by background jobs on Cloudflare.

Rebuild historyWeb archive + AI

Fly shops overwrite their reports weekly, so I recovered the history from web archive snapshots: 2,154 structured reports. Low-confidence or unattributed ones were dropped, not guessed.

ScoreRules, not AI

Fishability, hatch timing (adjusted for this year's degree-days) and river rankings are deterministic logic, so they can be tested. 648 automated tests.

Explaingpt-oss-120b on Groq

Writes the daily river brief and answers guide chat from the report archive. Responses are cached for 6 hours, so AI cost stays flat as traffic grows.

ShipCloudflare, Supabase, Stripe

React app on Cloudflare, accounts and data in Supabase, freemium premium tier through Stripe.

About 66K lines of code, built solo with Claude Code in 12 weeks (May 30 to Aug 20, 2026).

Step 05 / 05
Measured

How I measured it

Are the picks right?>90%accurate across 5+ rivers this summer

Do I use it myself?20+days on the water using it this summer

Can I trust the fly data?3AI judges must hit 90% approval on every fly pattern, with 3 retries, then human review

What I'd change

I built this to solve a problem I felt on the water and heard from the anglers closest to me. Because I loved the build and learning the tech, I kept adding features. Next time I'd narrow to the few that actually change a trip decision and launch with only those.