Thomas Austad Thomas-Austad

Talent Advisor Platform

A personal-use, ChatGPT-native AI career platform MVP.

Included

Evidence-grounded candidate profile agent
Compensation estimate agent
Greenhouse and Lever job connectors
Multi-factor job-match scoring
Truthful resume and cover-letter tailoring
Screening-question guardrails
Application approval state
DOCX resume export
FastAPI backend
MCP tool surface for ChatGPT
PostgreSQL starter schema

Run locally

Windows app launcher

Double-click scripts\\Start-TalentAdvisor.cmd. On first use, the graphical launcher confirms before installing private local application files, configures the local Ollama runtime, creates any missing private database settings, starts only loopback services, and opens the Career Workspace. It never submits an application; local approval and employer-page handoff remain separate, confirmed actions. See the Windows user guide for prerequisites and recovery steps.

If Docker Desktop is not ready or a local update is needed, the launcher shows a plain-language recovery message and does not reset existing data.

cp .env.example .env
# Ensure Ollama is running with qwen3:8b, then add a unique POSTGRES_PASSWORD.
# Set DATABASE_URL to the matching local PostgreSQL URL before running locally.
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\\Scripts\\activate
pip install -e .[dev]
docker compose up -d db
alembic upgrade head
uvicorn app.main:app --host 127.0.0.1 --reload

Health check: http://localhost:8000/health

Run MCP server:

cd widget
pnpm install --frozen-lockfile
pnpm build
cd ..
python -m app.mcp_server

MCP uses local stdio. Do not expose it through a development tunnel or public URL. The widget build is required because the local MCP resource serves the generated UI assets. Run pnpm test --run and pnpm typecheck from widget/ while developing the UI. The widget does not receive the local bearer credential and does not call the loopback API directly.

Typical workflow

POST /profiles
GET /profiles/{candidate_id}/review
PATCH /profiles/{candidate_id} to save user corrections
POST /jobs/search
POST /matches/{candidate_id}/{job_id}
POST /compensation/{candidate_id}
POST /applications/prepare
Resolve required user inputs
POST /applications/{application_id}/approve
Optionally review GET /applications/{application_id}/browser-handoff, then directly confirm POST /applications/{application_id}/browser-handoff to receive a validated employer-page link
Download /applications/{application_id}/resume.docx

Important boundary

This MVP prepares and tracks applications. Direct submission must be implemented only through approved ATS/job-board APIs or a user-controlled browser assistant. Never guess legally meaningful screening answers or submit without explicit approval.

Browser handoff is not application submission. It is available only for a locally approved package and a known HTTPS Greenhouse, Lever, or Ashby employer page. After direct confirmation, the app returns a user-activated external link without uploading materials, transferring employer credentials, prefilling a form, or representing a submission outcome.

API datetime fields are serialized as ISO 8601 timestamps with explicit UTC offsets. Sensitive screening questions are surfaced in unresolved_screening_questions, removed from generated screening answers, and block approval until resolved. Successful approvals and direct sensitive-screening confirmations write non-sensitive JSONL audit receipts to AUDIT_LOG_PATH (default var/audit/events.jsonl).

Local-only security model

This is a single-user, local-first application. The API is intended to bind only to a loopback address and requires a bearer credential for every private route. On a normal desktop installation, the application creates and stores that credential in the operating system credential store; LOCAL_ACCESS_TOKEN is a headless or automation fallback and must be at least 32 characters. The local operating-system user is the sole application principal, and no hosted accounts, sharing, or remote MCP transport are supported.

This model relies on the security of the user's operating-system account and disk encryption. It cannot protect information from an administrator, malware, or another process running as that same user.

Container safety

docker-compose.yml is for local development only. Its API and database ports bind to loopback so they are not exposed to the local network. Before using the stack, copy .env.example to .env, choose a unique local database password, and set DATABASE_URL to the matching local PostgreSQL connection URL.

Do not deploy this Compose file as a hosted service. This project is designed to run on the user's own computer with loopback-only access.

Development workflow

Repository-specific engineering rules live in AGENTS.md.

For explicit autonomous maintenance runs, use docs/autonomous-application-development-workflow.md. An example invocation and expected stop report are in docs/autonomous-workflow-example.md.

Navigation Menu