AI LinkedIn GANN

This project is a small experiment in GANN-style prompt optimization for tailoring a CV to a specific job description using LLMs.

Instead of training neural networks, it wires together three AI agents in a loop:

• Generator: given the job requirements, your CV, and a prompt/instruction, it produces text (e.g. selected most relevant work experiences plus summaries).
• Discriminator: given the same context plus the generator’s output, it scores how good the result is on a strict 0–10 scale. It gets an example of expected experiences to better judge the generator's ouput.
• Configurator: looks at the score, the current generator prompt, and the job/CV context and proposes a new, better prompt for the next round.

Running multiple rounds creates a simple Generative–Adversarial Neural Network (GANN)-like loop where the generator’s behaviour is steered, not by gradients, but by language feedback and prompt updates.

How the project is set up

• Entry point: src/index.ts

  • Loads three local text files from src/ai-inputs: cv.txt, job-requirements.txt, and expected-experiences.txt.
  • Runs a fixed max number of rounds (currently 5) of the GANN loop.
  • In each round it:
    • Builds a context-rich prompt from the job description and CV.
    • Calls the generator to produce output.
    • Feeds the result into the discriminator to get a score.
    • Asks the configurator for the next generator prompt if the score is below a target threshold.

• Generator agent: src/ai-agents/generator.ts

  • Thin wrapper around langchainClient.invoke with a simple system prompt (“You are a generator…”).
  • Accepts a free-form prompt and optional settings (temperature, max tokens, timeout).

• Discriminator agent: src/ai-agents/discriminator.ts

  • Uses a stricter system prompt and a zod schema to parse a numeric score between 0 and 10.
  • Receives job requirements, CV text, the generator instruction, and optionally expected experiences as a “ground truth” signal.

• Configurator agent: src/ai-agents/configurator.ts

  • Keeps a small in-memory history of previous prompts and scores.
  • Given the last prompt, config, score, and context, it suggests the next prompt that should lead to a higher score.
  • Uses a zod schema to validate the suggested prompt and returns plain text.

• Model client: src/lib/langchain-client.ts

  • Central place where model provider, model name, and invocation settings are configured.
  • Reads environment/API keys via dotenv and process.env (see .env.example).

The combination of these pieces gives you a minimal but expressive playground for experimenting with multi-agent prompt improvement loops on real CV and job text.

Getting Started

1. Install dependencies

   npm install

2. Configure your environment

   • Copy .env.example to .env:

     cp .env.example .env

   • Fill in any required API keys or model provider settings in .env (e.g. OpenAI or compatible endpoint).

3. Provide your data

   Ensure the following text files exist in src/ai-inputs:

   • cv.txt – your CV/resumé in plain text.
   • job-requirements.txt – pasted job description or requirements.
   • expected-experiences.txt – (optional but recommended) a few bullet points describing the ideal experiences that should be selected for this job. The discriminator uses this as ground truth.

4. Run the loop

   npx ts-node src/index.ts

   You’ll see, for each round:

   • The current generator prompt.
   • The generator output.
   • The discriminator score.
   • The configurator’s suggested next prompt.

5. Adjust and experiment

   • Tweak NUM_ROUNDS, the initial generator prompt, and GENERATOR_CONFIG in src/index.ts.
   • Modify the system prompts or model names in the agent files.
   • Change the scoring strictness or expected experiences to explore different optimisation behaviours.