AI Agent Engineer

AI Agent Engineer — Adzump

**Bangalore (Indiranagar / Koramangala area) · 5 days in-office · 3–6 yrs experience

About Adzump
Adzump is the agent that turns a 3-line chat prompt into a live Meta + Google ad campaign — creatives, targeting, budget, end to end. We're a small team inside Modlix, shipping to real advertisers, iterating weekly. Small enough that you'll touch every part of the system.

What you'll work on

• Build the AD Pilot agent — extend the chat-first flow: receipts, crafts, multi-logo handling, tool-text contracts. Real product work, not demos.
• Replace heuristics with agentic LLM decisions — push the system from hardcoded ladders to model-driven choices. Less if/else, more "let the model decide with the right context.
• Own behaviors end-to-end — from tool API design to prompt to eval. You debug your own hallucinations and ship the fix.

Must-have

• You've owned meaningful surface area of an LLM agent in production — tool API design, prompt structure, and eval setup. Doesn't have to be solo; does have to be yours to defend.
• You've built and iterated on an eval harness (offline + online) for an agent you worked on. This is the 2026 differentiator and we'll dig into it.
• Strong Python and comfort across the standard agent stack (any of OpenAI / Anthropic / Gemini SDKs in production).

Nice to have

• Multi-turn / stateful agent experience — memory, recovery from failures, human-in-the-loop
• TypeScript/React for occasional UI work on chat surfaces
• Willing to read Java service code occasionally for backend integrations

How we work
We default to small, direct solutions and add structure when it earns its keep. We plan before non-trivial changes, walk a 3–4 person panel through our reasoning, and update plans when someone makes a better point. We prefer agentic LLM decisions over heuristic ladders. We care about the user-facing voice of every tool.

Process

• 15-min intro call — context, mutual fit
• 30-min deep dive — your past agent work; we go specific
• Optional async (≤60 min) — code-walkthrough doc or short loom of an agent you've shipped, or a take-home if you prefer that format. If your best work is in private repos, this gives you a way to show it without re-interviewing under pressure.
• 30-min final — agentic system design + culture conversation
• End to end in 1 week.