🤗 Find the Pokemon you are w. PokéAPI, your resume & embeddings

This is a submission for Weekend Challenge: Passion Edition These two last weeks, my team mates started to to use Claude Code together with they yearly review : 📓 Versionner et builder l'eBook de son Entretien Annuel d'Evaluation sur Git(Hub) adriens Mar 25 '24 #markdown #pandoc #ebook #ai 2 reactions  comments to discover which Pokemon they are... and why. I found that really really fun... and started to wonder if I could automate that with only onPrem resources, with embeddings, ML... only with a simple laptop without GPU, a simple core i5 and 8 Gib. And of course, only with pure Open Source software ❣️ What you'll discover below is how I started to prototype it and make it happen. 🤗 Code The whole code source is available as a HF Space, see rastadidi/resume-to-pokemon for more... or to play wth it 🤓 To achieve this first prototype I : Used the data I already prepared with my registry.jsonresume.org/adriens Bundled dataset (built once). build_dataset.py fetches every species from the PokeAPI — name, types, base stats, sprite, genus and English Pokedex flavor text. For each Pokemon it also derives a professional-archetype profile from its types and stat spread (e.g. a Steel type → "a disciplined, precise, robust engineer of structured systems"), so career resumes and monster biology meet in the same trait vocabulary. Description + profile are embedded with BAAI/bge-m3 and committed as data/pokemon.json + data/embeddings.npy — so the app makes no PokeAPI calls at runtime. Resume → phrases. Sections that carry semantic signal — basics.summary, skills, work/volunteer, projects, interests — are each turned into a short phrase and embedded with the same model. (Administrative sections like education, certificates and languages are skipped.) Retrieve → rerank. Cosine similarity over the embeddings retrieves a shortlist of the closest Pokemon; a cross-encoder (BAAI/bge-reranker-v2-m3) then re-scores the (resume, Pokemon) pairs jointly for much sharper precision than cosine alone. The tool explains why by quoting the matched resume phrase and the Pokemon's own profile + Pokedex text. Ranking + relative fit. Pokemon are ranked by a blend of the rerank match and their base stats (adjustable). Because a broad resume matches many Pokemon similarly, raw scores cluster tightly and are unreadable — so the reported score is a relative fit: rerank scores are standardized across the shortlist and spread through a sigmoid, so the top clearly stands out (~100%) and the tail drops off. It's a fit relative to the candidate pool, not an absolute probability. The two best-fitting types are derived from the same shortlist, so they always agree with the ranked Pokemon. Calibrated confidence. Instead of a raw similarity number, a read-out reports how far the top match stands out from the field (a z-score over the shortlist), flagging decisive vs. diffuse, multi-type profiles. Type scores and the ranked Pokemon are derived from the same reranked shortlist, so the "best-fit typing" always agrees with the cards.