rogerhenley.dev
This site: a working system built in the open, argued for by running rather than by describing.
- My role
- Sole designer and developer
- Status
- In active development
- Since
- 2026
- Source
- github.com/reh3376/career-site
- Go
- Python
- Next.js
- Postgres
- pgvector
- Docker
- Caddy
The problem
A CV asserts. A reader has no way to check an assertion, so the ones that matter get discounted along with the ones that do not.
The approach
Build the claim instead of writing it. The site runs a job-description reviewer, a grounded assistant and a timed decision-quality experiment, all on one small server, with the specification, the architecture decisions and every diff public.
The reviewer reads a posting requirement by requirement against a corpus of real records and returns met, partial or unmet with the evidence it used. The score is arithmetic over those verdicts, computed in code: the model never emits a number.
Outcomes
Its own reviewer is graded by hand. 26 of 31 gradeable verdicts agree with mine, 83.9%, with two ungradeable set aside rather than counted as agreement. The target is 85%, so it is not there yet, and the number is published because it is not there yet.
The direction of the disagreements is the part worth knowing. All four hard disagreements were the reviewer refusing to credit work I can evidence. None were it crediting work I cannot. A tool its owner built to argue for him has an obvious failure mode, and so far it has failed the other way.
The specification is a living document with every requirement tracked to its state, and the gap between what it claims and what is built is published rather than tidied away.