CV agent: from manual to production
01 · SITUATION
The context
Adapting a consultant CV to a tender represented 30 to 45 minutes of manual work, repeated dozens of times per year, with heterogeneous output depending on who handled it. The real risk: losing a tender not on competence, but on how the dossier was presented.
02 · METHOD
4-step pipeline
Local document extraction
Raw data extraction from PDFs (pdfplumber). Local processing. No data leaves the infrastructure before validation.
AI constrained by strict schema
14 Pydantic fields, temperature 0. Two passes: extraction of the union of CV versions, then alignment with the tender specifications. [[To complete]] markers for gaps, [[To arbitrate]] for conflicts.
Deterministic decoupled rendering
PPTX generation in native python-pptx. Total AI/rendering decoupling = zero truncation, zero dependency on a third-party rendering service.
Mandatory human validation
The agent does not replace the consultant: it produces a structured, documented first draft. The final decision remains human.
ARCHITECTURE DIAGRAM (PIPELINE)
Orchestration pipeline.py · Entry points cli.py / server.py (CLI + FastAPI) · Local sandbox: no data leaves the infrastructure before validation.
USER JOURNEY (FROM FILE DROP TO FINAL CV)
LLM cost dominates total time (about 80% of the duration).
03 · RESULTS
In production
Next case
AI governance at large-group scale