Workybara is an AI job search engine I designed and built solo. It checks 210K+ company career sites and ranks 4.7M+ open jobs against the skills a person has actually used, instead of keyword overlap.
I own product, specs and evaluation, and direct Claude Code to write most of the code. Stack: SvelteKit, Supabase/Postgres, AWS, Qdrant vector search, and self-hosted open models via llama.cpp to keep inference cheap.
The interesting parts are below.
A fine-tuned 350M-parameter skill extractor that reads every job post (35% fewer hallucinations and 2.4x faster than the version before it, 99.8% valid JSON)
Custom embedding models for matching, now being replaced by a distilled 97M model with 16x less vector storage that matches unseen employers better
A blind three-judge LLM eval (DeepSeek, Claude, Gemini) over 4,131 jobs, used to cut wrongly matched skills by 44%
0 likes · 0 views · Nov 10, 2026
Solo-built AI job search engine that ranks 4.7M+ jobs from 210K+ career sites by real skills, using fine-tuned models I trained.
