jobops-copilot

JobOps Copilot is an AI-agent operations platform for the job search. It discovers and tracks opportunities in a CRM, then applies real LLMs, retrieval-augmented generation, and multi-step agents to analyze fit, research companies, prepare interviews, plan skill gaps, draft outreach, and surface time-series insights. The project explicitly positions itself as a responsible AI operations system rather than an auto-apply bot: it drafts and recommends, but never sends or fabricates. Key capabilities described in the README: - Multi-provider LLMs: a Python agent service routes to Anthropic Claude, Azure OpenAI, OpenAI, or Google Gemini via LangChain's init_chat_model, with structured-output validation. A deterministic mock fallback keeps the app working without keys. - Multi-step LangChain agents: interview-prep, company research (with a web-search tool), and skill-gap planning built on create_agent plus ToolStrategy. Outputs persist per job and can be regenerated. - In-app job discovery: ingests roles from Adzuna with a local keyword pre-rank, then computes an LLM fit score on open; includes recency filter, posted date, matched-skills snippet, and a scheduled discovery cron. - Add-job URL autofill: an SSRF-guarded fetch with a tiered extractor (JSON-LD JobPosting, OpenGraph, heuristic) behind an Autofill button. - Floating global assistant: a multi-turn, context-aware chat widget on every authenticated page, streamed end to end from Python through Express and Next.js, with sessionStorage persistence and accessibility support. - RAG over pgvector: resumes and job descriptions are embedded with Hugging Face sentence-transformers (PyTorch) and stored in Postgres pgvector; fit scoring is grounded in retrieved resume evidence. - Time-series telemetry intelligence: pandas trend, anomaly, and forecast analysis over the pipeline, narrated by an LLM, plus a synthetic EV battery telemetry demo. - UI: Next.js 16, React 19, Tailwind v4, shadcn/ui, light/dark themes, onboarding wizard, Clerk authentication, and Playwright verification across breakpoints. - Workflow automation: n8n webhooks for job intake, follow-up reminders, and weekly reports, with companion flows for Make.com and Zapier. - Production discipline: npm and Python CI (lint, typecheck, build, tests), protected main branch, Azure PostgreSQL, Blob Storage export, and App Service deploy workflow. - AI ops: Adzuna ingestion, Langfuse tracing, an eval harness (deterministic plus Ragas) with a two-tier CI gate, and runtime guardrails including per-user rate limiting, daily cost ceiling, contact-PII redaction, prompt-injection defense, and output moderation. Architecture: apps/web (Next.js) to apps/api (Express) to services/agent (Python/FastAPI, which owns the AI) to Azure PostgreSQL with pgvector, surrounded by Blob Storage, automation, and Azure platform services. The API delegates AI to the agent service when AGENT_SERVICE_URL is set, otherwise uses a mock. Database migrations are applied by the API at boot, and a readiness endpoint reports schema gaps. Tech stack: Next.js 16, React 19, TypeScript, Tailwind v4, shadcn/ui, Clerk; Express 5 with pg; Python 3.12, FastAPI, LangChain, sentence-transformers, pandas, psycopg/pgvector; Azure Database for PostgreSQL, Blob Storage, App Service, Container Apps; n8n, Make.com, and Zapier for automation. Safety and human approval: the system drafts outreach but never sends automatically, suggests resume emphasis but never fabricates experience, scores fit and researches companies while leaving decisions to the user, and keeps outputs structured, grounded, and auditable.