Description :

We're hiring an AI Engineer to own the core agentic system of our product : the tool-use loop, the prompts, the model routing, multi-agent delegation, and the evaluations that prove it all works. You'll be the person who decides which model runs which job, why, and how we measure it.

This is hands-on engineering, not research. You'll ship to real users.

What you'll actually do :

- Own the agent loop : Improve how the agent plans, calls tools, recovers from errors, manages context, and decides when to stop. Today the loop is a competent but vanilla tool-call loop there's a lot of room to make it smarter, faster, and cheaper.

- Tune the prompts : System prompts, tool descriptions, memory-retrieval prompts, planning prompts, and delegation prompts all need iteration. You'll own the prompt library and the discipline that keeps it from rotting.

- Design the model router : Cheap models for trivial turns, strong models for hard reasoning, vision models when needed, local models when privacy demands it. Each agent has a primary and a fallback. You'll design and ship a smarter routing layer.

- Improve tool calling : Tool registry, tool descriptions, argument schemas, error feedback, result shaping all affect agent reliability. You'll make tools more reliable and tool errors more recoverable.

- Make multi-agent delegation work : One agent handing a task to another sounds simple it isn't. Constraint passing, context scoping, result aggregation, deadlock avoidance, budget inheritance. You'll design the delegation protocol and prove it works on real tasks.

- Build the evaluation harness : Without evals, we're guessing. You'll build a test suite of real agent tasks (with golden outputs, automated graders, regression tracking) so every prompt change, model swap, or runtime tweak gets measured before it ships.

- Wrestle with memory : Retrieval quality is a major lever on agent performance. You'll improve how memory is written, summarized, indexed, and surfaced into prompts.

- Handle the unglamorous parts : Token budgets, cost tracking, prompt caching, context compaction, latency, retries, streaming. The platform succeeds or fails on these.

What we're looking for :

Required :

- 3+ years engineering experience, with at least 1+ year shipping LLM-powered products (not just prototypes production systems with real users)

- Deep familiarity with the Anthropic, OpenAI, or Google Gemini SDKs you've written real tool-use loops, not just chat completions

- Strong TypeScript / Node.js. You can read and modify a production codebase without hand-holding

- Genuine prompt-engineering instinct. You've debugged "why is the model doing X" and can articulate what you'd try and why

- Experience with agent frameworks or building one you understand tool-calling loops, ReAct, planning, context management

- Comfortable with eval design. You don't ship LLM changes without measuring them

- Pragmatic about model selection. You know when to use the big model and when not to

Strong bonus :

- Built or contributed to a multi-agent system (not just chained prompts)

- Experience with MCP (Model Context Protocol) tooling

- Familiarity with prompt caching, structured outputs, vision models, long-context handling

- Worked with local/open-weight models (vLLM, Ollama, llama.cpp)

- Track record reducing LLM cost and latency on a real product

- Built RAG / memory systems with real-world retrieval quality discipline (not just cosine_similarity > 0.7)

- Experience with browser automation agents, computer-use agents, or code-execution agents

- Comfortable doing the "writer" parts of prompt engineering you can write clear English instructions, not just engineer the scaffolding

Bonus we'd be excited about :

- You've published evaluations, prompts, or model comparisons others have used

- You've shipped agentic features that materially changed product metrics

- You've trained or fine-tuned models, even if you don't claim to be an ML researcher

What We Offer :

- Competitive salary.

- Opportunity to work with cutting-edge AI technology.

- A collaborative and inclusive work environment.