WordPress · AI matching
Corporate website and AI CV matching (Keoni AI)
A corporate website, plus a matching engine that ranks the CVs of a WordPress job board for each job offer, with an explanation for every candidate.
Kéoni Consulting needed a clear online presence and a faster way to shortlist candidates in a pool of more than 32,000 CVs. The goals: cut pre-selection time, improve the relevance of the profiles proposed, and keep control of the data by self-hosting the stack.
Simplified view of the matching flow and the deployment pipeline.
The first version (February to March 2026) chained a WordPress hook, an n8n workflow, an in-memory FAISS index and a write back to WordPress. In September 2026 I rewrote the engine around pgvector and text windows, with direct writes into WordPress. Since October the "Lancer IA" button calls the engine directly, and n8n is only used for extraction previews.
The first chain truncated text silently and kept results in memory. The new engine indexes CVs in overlapping text windows (no silent truncation), ranks a candidate on the best window rather than an average, and writes results itself.
Running the cross-encoder on all 33,000 CVs meant an offer never finished. It now re-reads the top 100 only, and every other CV keeps a real score from cosine similarity. Sending one result per CV in a single request was also replaced by writing the 300 best.
OCR was timing out on every page. Measurements showed one OCR took 1.1 s, but eight in parallel pushed the server load to 32. Tesseract is multi-threaded, so a wrapper limits it to one thread: the same eight now finish in 2.4 s, and PyTorch keeps its threads.
An audit showed empty or unreadable CVs ranking near the top, because semantic similarity saturates around 73%. A CV with no detected skill is now capped at 30%, generic skills weigh less, and one person gets one line even with several applications.
I answer within a day, usually faster.