Job hunting is broken. Between endlessly scrolling through job boards, manually tweaking cover letters, and hunting down recruiter emails, the process is exhausting. I realized I was spending more time formatting documents than actually building software.
I decided to fix it by treating myself as the "CEO" and building a digital workforce to do the heavy lifting. The result is Your Job Searcher, an autonomous, multi-agent orchestrator.
(Note: I wrote this article as part of my submission for the #AllThingsAgenticHackathon).
The Architecture: A "Blackboard" Memory System
Instead of writing a single, massive script that breaks easily, I built four distinct AI agents. To make them work together seamlessly, I used a Stateful Memory Pattern. Google Cloud Storage (GCS) acts as the central "brain." One agent writes its findings to the bucket, and the next agent reads from it.
Here is the breakdown of the workforce:
The Scout: Reads my live profile data and dynamically scrapes real Google Jobs and LinkedIn listings using SerpApi.
The Pitchman: Pulls the job leads from GCS and uses Gemini 3.5 Flash to generate a hyper-targeted cover letter without the usual AI fluff.
The Closer: Authenticates securely with the Gmail API via OAuth, drafts a subject line, and stages a live draft directly in my inbox.
The Reporter: A custom Gemma 270M model that audits the entire execution pipeline and saves a final markdown report back to the cloud.
The Tech Stack & Challenges
The frontend is built on Streamlit and deployed as a containerized web app on Google Cloud Run. Because security is paramount when dealing with OAuth tokens and APIs, I utilized Google Cloud Secret Manager to ensure zero hardcoded credentials existed in the codebase.
One of my biggest advantages in building this was my academic focus on English Linguistics. Understanding syntax, semantics, and language structures allowed me to engineer highly precise system prompts for the Gemini models, bridging the gap between human language and backend logic.
However, it wasn't without challenges. Adapting to the new Automatic Function Calling (AFC) requirements in the absolute latest google-genai Vertex AI SDK required a complete migration to the Chat interface (client.chats.create). Additionally, figuring out how to host the Gemma 270M model cost-effectively was a puzzle, but I successfully deployed it locally on a GCP e2-micro VM in the "Always Free" tier.
What I Learned
I previously built a web app on GCP, but I honestly felt Gemini Notebook (formerly NotebookLM) did the job better. With Your Job Searcher, however, I finally built something that feels incredibly practical and powerful.
The biggest lesson? AI is an incredible partner, but it still needs a CEO. You have to monitor its outputs, manage its state, and design resilient architectures like the GCS Blackboard to keep it on track.
