Werin Sourcing — AI-Augmented Recruitment Platform

Werin Group, a Moroccan recruitment firm, runs its sourcing through the BoondManager ATS, but two tasks stayed slow and unreliable when done by hand: classifying each CV by specialty domain and computing a candidate's real years of experience, then, for every open position, quickly surfacing the right candidates from a pool of several thousand CVs. Saad had to design an internal tool that could make this sourcing more reliable without replacing Boond or disrupting recruiters' habits, while working within the real constraints of shared hosting with no root or SSH access.

Under Saad's technical leadership, a team built Werin Sourcing, a platform made of several independent Node.js/Express services (a Boond gateway, a CV extractor, a sourcing engine) plus an Angular 19 frontend. The part that's operational today covers:

• automatic extraction of specialty domain (a fine-tuned classifier) and years of experience (a deterministic engine) from the raw CV text already stored in Boond, written back after review
• a background "bootstrap" job to recompute these fields across the entire existing candidate pool, with pause/resume, cost tracking, and an admin review screen before anything is written back
• for each open position in Boond, a hard filter by domain/experience followed by keyword-relevance ranking, complemented by an AI-generated criteria grid extracted from the free-text job description

The product's later stages — semantic embedding search, outreach automation, candidate-dossier generation — are documented on the roadmap but not yet built.

A dated code-and-spec audit directly contradicted the project's own internal tracker, which claimed 95% completion on phase one: the audit found that a module marked "done" actually meant "coded in isolation," not "wired together and working end-to-end," and catalogued roughly ten concrete bugs breaking the pipeline. Rather than ignore that finding, the team documented and fixed it. The real deployment to shared hosting produced its own incident report detailing 9 concrete production bugs (dependency conflicts, schema mismatches, double URL-prefixing, an authentication bypass), all fixed and documented — a discipline of systematic verification rather than a silent go-live.

Computing years of experience originally relied on a local LLM (Ollama, Qwen2.5-3B), abandoned after repeated failures under load and clearly identified reasoning errors (a missing end date interpreted as "today," adding phantom years; internships counted as work experience). It was replaced with a deterministic date-parsing engine, hundreds of times faster and fully reproducible. On the classification side, rather than merging similar categories, the platform lets a CV carry multiple domain labels at once, and deliberately silences low-confidence classes into a manual review queue instead of guessing. The target architecture called for Redis, Qdrant and local VPS inference; the real constraints of shared hosting led the team to explicitly drop them for the version currently deployed.