CV Creator Agent

A lightweight LLM-powered pipeline to tailor a master CV to a target Job Description.

This repository provides a small workflow of LLM-driven processing steps (called "agents") that analyze a job description, draft a concise professional summary, and rewrite experience bullets to align with the target role and ATS keywords.

Key Concepts

• Workflow (pipeline) — not an autonomous planning agent. Orchestration is explicit and deterministic via a StateGraph.
• Agents — independent processing nodes that accept the shared state, call an LLM with a prompt, parse outputs, and return updates to the state.
• State — a TypedDict (src/state.py) representing inputs, analysis, generated outputs, and errors.

Features

• Extracts tech and soft keywords and role focus from a Job Description.
• Generates a 3-sentence professional summary tuned to the JD.
• Rewrites experience bullets for three roles using an ATS-friendly prompt.
• Tests to validate basic pipeline runs (no external orchestration required).

Tech Stack

• Python 3.10+
• LangChain-style primitives (langchain_core) for prompts and parsers
• Providers: langchain_google_genai and langchain_ollama (LLM clients) — configured in src/utils.py
• Orchestration: langgraph (StateGraph) at src/graph.py
• Tests: pytest (see tests/)

Repository Layout (important files)

• src/graph.py — the orchestrator: builds the StateGraph and compiles app.
• src/agents/analyst.py — extracts keywords and role focus from JD.
• src/agents/summary.py — drafts the professional summary.
• src/agents/experience.py — rewrites experience bullets using an ATS prompt.
• src/agents/prompts/experience.toml — canonical experience prompt template.
• src/utils.py — LLM client factory and helper functions.
• src/state.py — CVState TypedDict describing the workflow state.
• tests/ — integration-style tests that invoke app.invoke(initial_state).

Getting Started

Prerequisites

• Add any LLM provider keys to a .env file in the repo root (see src/agents/__init__.py which loads .env). Common vars:
  • GOOGLE_API_KEY (for langchain_google_genai)
  • Any Ollama/local model host configuration if applicable

Install dependencies (example using pip):

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Note: This repo uses provider-specific packages (langchain_google_genai, langchain_ollama, langchain_core, langgraph). If you rely on a pyproject.toml, install via that (pip install -e .) or adapt to your environment.

Running tests

# from repo root
python -m pytest -q

Quick usage (interactive)

from src.graph import app
from src.utils import load_master_data

initial_state = {
    "jd_text": "<PASTE JOB DESCRIPTION>",
    "master_cv": load_master_data(),
    "analysis": {},
    "summary": "",
    "experience_bullets": {},
    "selected_project_ids": [],
    "project_bullets": {},
    "skills": {},
    "review": {}
}

output = app.invoke(initial_state)
print(output["analysis"])   # analyst output
print(output["summary"])    # generated summary
print(output["experience_bullets"])  # tailored bullets

Design Notes

• The project intentionally implements a pipeline: each node is responsible for a single transformation and returns only the portion of state it owns. The compiled app runs the nodes in sequence and stops on error.
• Each agent uses a prompt-driven pattern: prepare data → call LLM → parse output → return state diff.
• src/agents/prompts/experience.toml contains a stricter, production-ready prompt for the experience rewriting step.

Recommendations / Next Steps

• Harden LLM calls with timeouts, retries, and backoff (wrap SDK calls with tenacity or provider-native retry configs).
• Add a circuit-breaker or rate-limiter to avoid hammering providers.
• Add an orchestration entrypoint (CLI or HTTP worker) to queue runs, persist intermediate state, and resume failures.
• If you want an autonomous agent (planner + tools), provide a toolset (search, retriever, file access) and an action interpreter; otherwise keep the clean pipeline model for reproducibility.

Contributing

• Open issues or PRs; keep prompts and state shape explicit.

License

MIT