Career Copilot

Career Copilot

A multi-agent CV × JD pipeline on Google ADK v2.

Live demo · Project write-up

---

Career Copilot reads a CV against a job description and produces a different output depending on who is asking. Recruiters get a fit verdict and a LinkedIn outreach draft that quotes a real achievement from the CV. Candidates get a company brief researched live through Tavily MCP and a tailored interview-prep bundle.

The point of the project is the orchestration. Google ADK v2 shipped its graph workflow API recently, in the same spirit as LangGraph. This is an end-to-end test of it on a real use case.

Why a graph

A single big-prompt agent works until you need any of: deterministic branching, parallel work, typed contracts at every step, or per-step tracing. Once you do, a graph stops being optional.

• Explicit topology: nodes, edges, parallel branches, conditional routing, and sync points are first-class.
• Testable units: each node can be tested, swapped, or traced in isolation.
• Typed state: contracts at every boundary instead of free-form prompt parsing.
• Predictable cost: you know upfront which steps run, in what order, and which model serves each.

How it works

flowchart LR
    IN(["CV + JD+ mode"]):::io

    IN --> CV[CV Parser]:::agent
    IN --> JD[JD Parser]:::agent
    CV --> MR(["Mode Router"]):::router
    JD --> MR

    MR -- recruiter --> FA[Fit Analyzer]:::agent
    FA --> VR(["Verdict Router"]):::router
    VR -- fit / borderline --> OW["Outreach Writermedium"]:::agent
    VR -- no_fit --> GE[Gap Explainer]:::agent

    MR -- candidate --> CFORK((·)):::fork
    CFORK --> RA["Research Agent+ Tavily MCP"]:::agent
    CFORK --> CO[CV Optimizer]:::agent
    RA --> IP[Interview Prep]:::agent
    CO --> IP

    OW --> OUT1(["RecruiterFit"]):::io
    GE --> OUT2(["RecruiterNoFit"]):::io
    IP --> OUT3(["CandidateResp"]):::io

    classDef agent fill:#E3F2FD,stroke:#1565C0,stroke-width:2px,color:#0D47A1
    classDef router fill:#E8F5E9,stroke:#2E7D32,stroke-width:1.5px,color:#1B5E20
    classDef fork fill:#9E9E9E,stroke:#616161,stroke-width:1px,color:#9E9E9E
    classDef io fill:#F5F5F5,stroke:#616161,stroke-width:1px,color:#212121

CV and JD are parsed in parallel. The mode router (a FunctionNode, not an LLM) splits the flow:

• Recruiter branch: linear chain. Fit Analyzer scores fit, Verdict Router routes to Outreach Writer (on fit / borderline) or Gap Explainer (on no_fit).
• Candidate branch: parallel fan-out. Research Agent (Tavily MCP) and CV Optimizer run concurrently, synchronize through a JoinNode, then feed Interview Prep.

Internal JoinNodes are omitted from the diagram for clarity. See [app/agent/agent.py](backend/app/agent/agent.py) for the full topology.

Architecture decisions

The graph encodes a deliberate separation between deterministic control flow and LLM-driven generation. Routing and synchronization stay in pure Python; only generation crosses an LLM boundary, and every output is constrained by a Pydantic schema before it leaves a node.

• Control flow stays out of the LLMs. Routing lives in FunctionNodes that decide branches in pure Python on already-typed state. JoinNodes wait for parallel branches to complete before firing downstream. No LLM ever picks which branch fires next, which removes a whole class of failure modes from the critical path.
• One schema per agent. Each LLM agent declares its own output_schema. Invalid output fails the node loud and fast instead of corrupting state in a downstream agent.
• Tool-use escape hatch. ADK currently disallows combining output_schema with tools on gpt-5.4-mini. The Research Agent uses the Tavily MCP toolset and emits CompanyIntelligence as a JSON string, validated with model_validate_json at the API boundary so the typed-state invariant is preserved at the edge.

Agents

All agents run OpenAI gpt-5.4-mini via LiteLlm. The Outreach Writer is the only one dialled up to medium reasoning effort.

| Agent | Role | Output schema | | ------------------ | --------- | ------------------------- | | CV Parser | Parser | ParsedCV | | JD Parser | Parser | ParsedJD | | Fit Analyzer | Recruiter | FitVerdict | | Outreach Writer | Recruiter | OutreachDraft | | Gap Explainer | Recruiter | GapReport | | Research Agent | Candidate | CompanyIntelligence | | CV Optimizer | Candidate | CVOptimizationBundle | | Interview Prep | Candidate | InterviewPrepBundle |

Tracing

Every /v1/analyze run is traced end-to-end with Langfuse. Sub-agent calls, tool calls, latency, and token counts nest under a parent agent observation. The filterable trace attributes propagated to Langfuse carry only metadata and sizes (mode, model, version, cv_chars, jd_chars); raw CV and JD content is kept out of the trace tags so traces stay free of PII.

Stack

| Layer | Stack | | --------- | ------------------------------------------------------------------ | | Backend | Python 3.12, FastAPI, Google ADK v2, uv | | Models | OpenAI gpt-5.4-mini via LiteLlm | | Research | Tavily via MCP (McpToolset) | | Frontend | Next.js 15, React 19, Tailwind v4, shadcn/ui, TanStack Query, Zod | | Tracing | Langfuse | | Deploy | Docker (Cloud Run / Fly.io / HuggingFace Spaces) |

Run it

Prerequisites

• Python 3.12 with uv
• An OpenAI API key for gpt-5.4-mini
• A Tavily API key for the candidate-mode Research Agent

Install and configure

cd backend
uv sync
cp .env.example .env   # fill in OPENAI_API_KEY and TAVILY_API_KEY

Start the backend

uv run uvicorn app.main:app --reload --port 8080

API docs at . Try POST /v1/analyze with {"cv_text": "...", "jd_text": "...", "mode": "recruiter" | "candidate"}.

Run the tests

uv run pytest tests/ -q

License

[MIT](./LICENSE)

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Naji-Najari
• Source: Naji-Najari/career-copilot
• License: MIT
• Homepage: https://career-copilot.najinajari.com

Install and usage instructions live in the source repository linked above.