Athena Onboard

AI-generated, HR-reviewed onboarding plans on a serverless AWS stack.

HR enters a new joiner's employee ID → Athena fetches their profile → picks the most relevant training with embeddings and cosine similarity → drafts a 3-week plan with an LLM → HR reviews and approves → the plan is emailed to the new joiner, with every step in an audit trail.

This is a public demo rebuild of a hackathon project I worked on. The original ran on Amazon Bedrock (Titan Text + Titan Embeddings); this version uses the Google Gemini API free tier so the demo costs nothing. All employees and training content are synthetic.

Architecture

Static UI (GitHub Pages)
   │
   ▼
API Gateway (HTTP API, throttled)
   ├── GET  /employees/{id}      → get-employee-details  (mock HR system)
   ├── POST /plans               → generate-plan         (cosine similarity + LLM → DynamoDB)
   ├── GET  /plans/{id}          → review-plan
   ├── PUT  /plans/{id}          → review-plan           (HR approve / reject, versioning, audit)
   └── POST /plans/{id}/send     → send-email            (SES, or an email preview in demo mode)

Piece	Choice
Compute	AWS Lambda (Python 3.12, arm64)
API	API Gateway HTTP API with throttling
Storage	DynamoDB (plans, versions, audit trail)
Email	Amazon SES (optional; preview mode when not configured)
LLM	Gemini (gemini-2.5-flash) for plan generation, gemini-embedding-001 for content selection
Fallback	Template planner + bag-of-words cosine when no API key is set or the free quota is used up
UI	Plain HTML/CSS/JS single page; runs fully in the browser in demo mode
IaC	AWS SAM (backend/template.yaml)

Guarding against hallucinations: the LLM only sees the top training modules picked by cosine similarity and must reference them by ID; anything else is dropped before the plan is saved.

Run the UI locally (demo mode)

Demo mode needs no backend: config.js has an empty API URL, so mock-api.js stands in for the Lambdas.

mkdir -p site/data && cp -r frontend/* site/ && cp backend/src/data/*.json site/data/
python -m http.server 8000 --directory site

Open http://localhost:8000 and try EMP1001.

Deploy the backend

Prerequisites: an AWS account, the AWS CLI (aws configure) and the AWS SAM CLI. Optional: a free Gemini API key from Google AI Studio.

cd backend
sam build
sam deploy --guided   # stack name: athena-onboard; paste the Gemini key when asked (or leave empty)

Then put the ApiUrl output into frontend/config.js and push; GitHub Actions republishes the UI.

Cost: Lambda and DynamoDB stay within the AWS free tier at demo traffic; the API is throttled and plan generation is limited to 2 concurrent runs. Set an AWS budget alert anyway.

Project layout

backend/
  template.yaml              SAM template (API, Lambdas, DynamoDB)
  src/
    get_employee_details.py  GET /employees/{id}
    generate_plan.py         POST /plans
    review_plan.py           GET / PUT /plans/{id}
    send_email.py            POST /plans/{id}/send
    llm.py                   Gemini text + embeddings, cosine similarity, offline fallback
    common.py                responses, data loading, DynamoDB + audit helpers
    data/                    synthetic employees and training catalogue
frontend/                    single-page UI (index.html, app.js, mock-api.js, styles.css, config.js)