Website:
blynkads.com
Job details:
About the role
We're building a chat-based assistant that lets advertisers optimize their Meta campaigns in plain language. Users ask open-ended questions like "I want more leads," "my leads are coming from too far away," or "my leads aren't answering." The system has to work out what's actually wrong, pull the right campaign data, and recommend or take action.
You'll design and build the agentic core of this product. It's an active, in-development project, so you'll be joining a live codebase, not starting from a blank page.
What you'll do
- Design and build a multi-step agent that turns vague user goals into diagnosis and action on Meta campaigns
- Implement tool/function calling against our FastAPI backend and the Meta Marketing API
- Build agent workflows with LangGraph (state, branching, retries, human-in-the-loop checkpoints)
- Enforce structured outputs (JSON schema / Pydantic) so agent actions are safe and predictable
- Add guardrails so the agent never takes a costly campaign action without validation
- Set up evaluations and tracing (e.g. LangSmith or Langfuse) to measure accuracy, latency, and cost per conversation
- Containerize and deploy services on AWS (EKS, EC2) with Docker
- Work closely with the frontend team (Next.js) on streaming chat responses and agent state in the UI
Must have
- 2+ years of hands-on Python development, including async code and API design with FastAPI
- At least one agentic system you have built and shipped (not just a tutorial or wrapper around a single prompt)
- Practical experience with LLM tool/function calling and multi-step agent patterns (ReAct, planner–executor, or graph-based workflows)
- Working knowledge of LangGraph or an equivalent orchestration framework
- Experience evaluating LLM outputs: test sets, failure analysis, prompt iteration based on results
- Docker, AWS (EKS, EC2), Linux servers and SSH access
- Git and GitHub workflows (branches, PRs, code reviews)
- Clear written and spoken communication in English. You'll explain agent behaviour and trade-offs to non-technical teammates.
Good to have
- Experience with the Meta Marketing API or any ad-tech/performance-marketing product
- RAG or vector search (pgvector, Qdrant, Pinecone) for grounding agent decisions
- Redis, PostgreSQL, or MongoDB
- CI/CD pipelines (GitHub Actions)
- Basic familiarity with Next.js
How to apply
Send your resume, your GitHub profile, and a short note (5–6 lines) describing one agent you've built: what it did, what broke, and how you fixed it. Applications without a GitHub link or project description won't be reviewed.
Click on Apply to know more.
