Agentic AI Workflow for Scalable Document Generation

Overview

Designed a nine-agent LangGraph state machine with specialised nodes for job-description acquisition, structured JD analysis (skills decomposition, keyword extraction and seniority signals), company research with web search, comparative fit assessment, document generation, claims-map validation, a QA audit with iterative revision, LaTeX rendering, and versioned Google Drive storage. Agents exchange structured JSON contracts: the JD analysis agent produces a schema that downstream generators consume as a tailoring checklist, which lifted first-pass audit scores and cut wasted revision cycles.

Engineered production-grade reliability: thread-safe parallel processing with per-thread API clients and SSL serialisation, exponential backoff on transient errors, idempotent processing with versioned outputs, and observability through per-row cost tracking, processing-time logging and email failure alerts. Claims-map validation acts as a hard truthfulness constraint, cross-referencing every generated claim against verified source data and stripping anything invented.

Tuned for cost and throughput with tiered model assignment, company-research caching, early exit on audit-score plateaus and ThreadPoolExecutor parallelism, reducing estimated per-row cost by about 50% and batch time from 37 hours to under 8 hours for 200 applications. Integrated Google Workspace APIs (Sheets, Drive, Docs) for input, versioned output and traceability.