PathOS

Inspiration

Job seekers often confuse “I can do this job” with “this job is worth pursuing.” A role can match a CV while offering weak long-term career value, unclear company quality, or too much delivery and operational work. PathOS was built to make those trade-offs visible and evidence-based.

What it does

PathOS turns a candidate CV and a job description into an explainable career decision. The current workflow:

1. Upload a CV and review an evidence-backed Career Profile.
2. Paste or upload a text-based PDF job description.
3. Extract the role, mandatory and preferred requirements, responsibilities, and delivery warnings.
4. Evaluate candidate fit, opportunity quality, and career-path alignment separately.
5. Show evidence, gaps, unknowns, citations, and one of three recommendations: Strongly Recommended, Explore Further, or Not Recommended.
6. Track the opportunity while keeping the system recommendation separate from the user’s own decision.

PathOS does not submit applications, contact recruiters, or invent candidate achievements. Material profile claims remain linked to source evidence, and protected state changes require explicit user confirmation.

How we built it

PathOS is a Google ADK multi-agent application behind FastAPI. A user-facing Coordinator delegates bounded work to a Career Profile Agent, Job Intake Specialist, Evaluation Supervisor, Fit Assessor, Quality Researcher, Path Alignment Analyst, and Delivery Detector. Deterministic Python services validate structured outputs and own protected database writes.

The deployed stack uses:

• Gemini 3.6 Flash through Vertex AI
• Google Agent Development Kit (ADK) and Google GenAI SDK
• Cloud Run for the FastAPI application, browser UI, and ADK/A2A routes
• Cloud SQL for PostgreSQL 16 with pgvector
• Vertex AI Google Search grounding for cited company research
• Secret Manager for database configuration
• Cloud Logging and Cloud Trace for observability
• Cloud Storage and BigQuery for telemetry infrastructure
• Terraform for reproducible Google Cloud infrastructure

Challenges we ran into

The hardest challenge was preserving evidence and user authority while coordinating multiple agents. CVs, job descriptions, and retrieved pages are untrusted inputs, so specialist agents remain read-only and deterministic services validate important state transitions.

We also had to separate three questions that many matching tools collapse into one score: Can the candidate perform the role? Is the opportunity attractive? Does it support the candidate’s long-term direction? PathOS keeps these assessments separate and exposes missing information instead of hiding uncertainty.

Deploying the complete workflow required packaging the UI, ADK runtime, migrations, Cloud SQL connectivity, secrets, search grounding, and logging into a reproducible Cloud Run deployment.

Accomplishments that we're proud of

• A working public Google Cloud deployment of the complete CV-to-job-decision workflow
• Evidence-linked candidate profiles and requirement matching
• A coordinated ADK agent hierarchy with bounded responsibilities
• Cited company and industry research through Vertex AI search grounding
• Qualitative recommendations that expose drivers, unknowns, and conflicts
• Append-only job tracking that preserves the distinction between system advice and user decisions
• Reproducible infrastructure, migrations, and local spin-up instructions

What we learned

Multi-agent design is most valuable when agents need different context, authority, and evaluation criteria—not simply because multiple agents sound impressive. We also learned that deterministic validation and provenance are essential for trustworthy personal agents. The model can analyze and propose, but evidence, permissions, versioning, and protected writes need explicit system boundaries.

Current limitations

This submission focuses on one stable decision workflow. Public job-page retrieval fallback, full profile correction UI, protected evaluation reruns, durable asynchronous operation polling, CV tailoring, Gmail ingestion, interview coaching, and application submission are intentionally outside the demonstrated scope.

Pre-existing and third-party code

PathOS was created during the submission period using the Google agents-cli ADK scaffold. It uses Google ADK, A2A SDK, Google GenAI/Vertex AI SDKs, FastAPI, Psycopg, pgvector, pypdf, python-docx, Uvicorn, OpenTelemetry, Terraform Google providers, and their open-source dependencies. The PathOS-specific agents, services, migrations, tests, UI, and product logic were built for this project.

Built With

• cloud-logging
• cloud-run
• cloud-sql
• cloud-trace
• fastapi
• gemini-3.6-flash
• google-adk
• google-genai-sdk
• pgvector
• postgresql
• python
• secret-manager
• terraform
• vertex-ai