AARLP – AI-Agentic Recruitment Lifecycle Platform

Inspiration

Hiring is broken. Recruiters spend hours writing job descriptions, manually screening resumes, and scheduling calls — only to repeat the same process for every role. We wanted to build something that handles the entire recruitment lifecycle autonomously, with a human only stepping in at key decision points. The Amazon Nova hackathon gave us the perfect reason to make it real.

What It Does

AARLP (AI-Agentic Recruitment Lifecycle Platform) automates hiring end-to-end using Amazon Nova on AWS Bedrock:

• JD Generation — A recruiter fills a simple form. Nova 2 Lite drafts a full, SEO-optimized Job Description in seconds.
• Human-in-the-Loop Editing — The recruiter reviews, edits inline, or gives natural-language feedback for AI regeneration.
• Semantic Candidate Ranking — Resumes are embedded using Nova 2 Multimodal Embeddings and ranked by cosine similarity against the JD vector — no keyword matching.
• Voice Prescreening (Architecture Ready) — Nova 2 Sonic is wired as a voice provider for automated phone screenings.
• LangGraph Agentic Workflow — The entire pipeline is a stateful LangGraph graph with checkpointing, so the agent can pause, wait for human approval, and resume.

How We Built It

The AI provider is fully swappable via a single env var (AI_PROVIDER=bedrock). All JD generation, regeneration, prescreening question creation, and candidate response scoring routes through invoke_nova_model() → Bedrock runtime → Nova 2 Lite.

Challenges We Faced

• Nova response format — Nova's Bedrock response uses a nested output.message.content[].text structure, unlike OpenAI's flat response. We built a dedicated bedrock_utils.py with proper message formatting and response parsing to handle this cleanly.
• Async Bedrock client — boto3 is synchronous by default. We integrated aioboto3 with async context managers so every Nova call is non-blocking inside FastAPI's async routes.
• Embedding dimension mismatch — Nova 2 Multimodal Embeddings produce 1024-dim vectors (not 1536 like OpenAI). We ensured Pinecone index creation, storage, and similarity scoring all respect the correct dimension dynamically based on the active provider.
• Stateful agentic workflow — Designing a LangGraph graph that correctly pauses for human approval (JD review, shortlist approval) without losing state across long-running recruitment pipelines required careful checkpoint design.

What We Learned

• Amazon Nova 2 Lite is fast and highly capable for structured JSON generation tasks like JD writing — with proper system prompting it produces schema-valid output consistently.
• Nova 2 Multimodal Embeddings open up rich cross-modal semantic search possibilities well beyond text-only matching.
• LangGraph's checkpointing + human-in-the-loop pattern is a powerful pattern for real-world agentic applications where full automation isn't always desirable.

Built With

• aioboto3
• amazon-web-services
• awsnovasonic
• bedrock
• css
• embeddings
• fastapi
• langgraph
• multimodal
• next.js
• pinecone
• postgresql
• pydantic
• python
• redis
• tailwind
• typescript