Career-Council

CareerCouncil reimagines the job application experience as a live council of five specialized AI agents – each with a distinct personality, expertise, and role – that collaborate in real time to transform a job description into a tailored, ATS-optimized LaTeX resume. Rather than a single chatbot answering questions, users watch a multi-agent deliberation unfold: Brian routes intent, Rita scrapes live company intelligence via Serper + Firecrawl, Marcus frames career narrative strategy, Robin archaeologically extracts verbatim resume evidence, and Ryan architects the final LaTeX output. Every agent is visible simultaneously with individual typing indicators, colored avatars, and streamed responses – making the AI’s reasoning process itself the UI, not just the output.The interface shatters the standard chat convention by treating the conversation as a visual pipeline graph – users see parallel agent execution (Rita and Marcus run simultaneously), staged handoffs (Robin waits for both before extracting evidence), and real-time status for each node in the chain. The sidebar tracks every job application as a cached artifact, enabling instant cache hits on repeat applications. Resume upload is a silent background operation – no form submission, no page reload – with a live pill indicator confirming the resume is “armed” for the next JD paste.Stability is achieved through a layered resilience architecture: exponential backoff retry on Anthropic 529 overloads, per-agent model selection (Haiku for fast extraction, Opus for quality generation), Redis-backed session and job artifact caching with TTL management, WebSocket reconnection logic, and graceful fallback messaging at every pipeline stage so a single agent failure never kills the session.
CareerCouncil – Project Overview
Description
CareerCouncil reimagines the job application experience as a live council of five specialized AI agents – each with a distinct personality, expertise, and role – that collaborate in real time to transform a job description into a tailored, ATS-optimized LaTeX resume. Rather than a single chatbot answering questions, users watch a multi-agent deliberation unfold: Brian routes intent, Rita scrapes live company intelligence via Serper + Firecrawl, Marcus frames career narrative strategy, Robin archaeologically extracts verbatim resume evidence, and Ryan architects the final LaTeX output. Every agent is visible simultaneously with individual typing indicators, colored avatars, and streamed responses – making the AI’s reasoning process itself the UI, not just the output.
The interface shatters the standard chat convention by treating the conversation as a visual pipeline graph – users see parallel agent execution (Rita and Marcus run simultaneously), staged handoffs (Robin waits for both before extracting evidence), and real-time status for each node in the chain. The sidebar tracks every job application as a cached artifact, enabling instant cache hits on repeat applications. Resume upload is a silent background operation – no form submission, no page reload – with a live pill indicator confirming the resume is “armed” for the next JD paste.
Stability is achieved through a layered resilience architecture: exponential backoff retry on Anthropic 529 overloads, per-agent model selection (Haiku for fast extraction, Opus for quality generation), Redis-backed session and job artifact caching with TTL management, WebSocket reconnection logic, and graceful fallback messaging at every pipeline stage so a single agent failure never kills the session.

Tech Stack
Backend

Python 3.11 + FastAPI – async WebSocket server, REST endpoints
Anthropic Python SDK – multi-model agent calls (claude-haiku-4-5, claude-sonnet-4-6, claude-opus-4-6)
Redis (via redis[hiredis]) – session state, resume storage, job artifact caching, cache hit detection
Serper API – real-time Google search for company intelligence (Rita)
Firecrawl API – structured web scraping of company/job pages (Rita)
python-docx – server-side resume parsing
Uvicorn – ASGI server with hot reload

Frontend

React 18 + Vite – component-based UI with fast HMR
Tailwind CSS – utility-first styling, dark theme
mammoth.js – client-side .docx → plain text parsing (zero server upload for common case)
react-markdown + remark-gfm – rich markdown rendering in agent bubbles
Native WebSocket API – persistent bidirectional connection with reconnect logic

Infrastructure / DX

Render.com – backend hosting with managed Redis
Vercel – frontend CDN deployment
GitHub – version control, Render CI/CD trigger

Architecture pattern: staged multi-agent pipeline with parallel execution lanes, Redis-backed artifact store, and model-tier routing – Haiku for speed-critical extraction, Opus for quality-critical generation.

Continuation of: https://www.forge-your-future.com
I’m making it multi-agentic, with agent-to-agent interactions!