Parallel invention: Kestrel vs career-ops
In February 2026 I started building an AI job-search platform for my own search. On April 4, santifer's career-ops went public. On April 8, so did my Kestrel. Four days apart, built independently -- neither of us knew the other existed. Theirs is at 62,000+ stars (62,874 when I checked on 2026-08-05). Mine has 5. This page is a teardown of that gap, because the gap isn't where people assume.
The timeline
| Date | Kestrel (me) | career-ops (santifer) |
|---|---|---|
| Feb 2026 | private build starts | -- |
| 2026-04-04 | -- | public repo created |
| 2026-04-08 | public repo created | -- |
| Apr-Jun 2026 | geo gate, provider chain, state machine, eval harness | Product Hunt, WIRED, Business Insider, 16 README translations, plugin registry |
| Jul 2026 | geo classifier v2 live: 93.6% recall on a 277-item blind set | 61,746 stars (2026-07-27) |
Nobody "found it first." The idea didn't meaningfully exist to find until April, by which point both systems were already built. Two people surfed the same wave; one of them also built a boat ramp.
What they did better (it wasn't the code)
I read their repo properly before writing this, and the honest verdict is admiration. Four deliberate launch decisions made the difference:
- Install friction. career-ops is skills you drop into the CLI you already run, with 4 npm dependencies. Kestrel is a venv, Docker, React, database migrations, and a multi-provider config matrix. Friction is the top-of-funnel of open source, and I lost that funnel before anyone saw a feature.
- Story-first packaging. Their README is a narrative with a protagonist ("I spent months applying the hard way, so I engineered the system I wish I had"), with their own usage numbers up front. Kestrel's README is a feature list. Feature lists don't get retold.
- Manufactured distribution. Product Hunt launch, press outreach that landed WIRED and Business Insider, translations, Discord, a contributor ladder, welcome bots. None of this is code. All of it is engineering.
- The repo WAS the job search. santifer pointed the artifact outward, and it worked: the repo landed its author a Head of AI role. I pointed my tool inward, at my own pipeline. (Which was the right priority call for getting hired, and I'd make it again. But it explains the star count better than any line of code does.)
What the small repo has
Same idea, different jobs. Their build is optimized for adoption; mine is optimized for being right, because I run my own search through it and a wrong answer costs me interviews. Where the production version is ahead: a geo/work-eligibility engine measured at 93.6% recall / 74.6% precision on a 277-item human-judged blind set (their location handling treats location as a scan string); eval gates in CI rather than a small golden set; provider preflight and fallback hardened by actual incidents; and a real database with a state machine underneath, where theirs keeps state in markdown. Full detail is in the Kestrel case study.
One cautionary data point before anyone (me included) gets star envy: AIHawk, the previous generation of this same idea, peaked around 30k stars and was archived by its own maintainers within 21 months. Stars are a wave-height reading, not a moat.
What I take from it
Distribution is an engineering discipline. It has friction budgets, and it deserves the same rigor as the geo gate got. I treated launch as an afterthought; santifer treated it as the product. Both of us were right about the technology at the same time; the difference was almost all launch craft.
The line I use in interviews: I independently built the category; theirs is the famous one, mine is the production version with measured evals. The next thing I launch gets a launch plan. This one got a job search.