Senior AI Engineer — Inference & Agent Systems

Arcana Analytics

About The Position

Arcana is building AI agents that synthesize information across heterogeneous sources and deliver structured, reasoned answers in real time. The product only works if the agents are fast, reliable, and correct — not approximately correct. Our stack: Go + Temporal for orchestration, a Plan-Execute-Synthesize agent architecture, and an evaluation harness we use to measure every regression. The problems are hard. The latency bar is aggressive. The accuracy requirements are unforgiving. The Work Inference Optimization - Drive TTFT below 400ms for multi-step agent pipelines - Streaming optimization — first token to user while sub-agents are still running - KV cache strategy, prompt compression, dynamic context window management - Multi-provider routing: model selection by latency, cost, and task type across OpenAI, Anthropic, Gemini, and open-weight models Agent Architecture - Design and implement Plan-Execute-Synthesize pipelines that run sub-agents in parallel DAGs, not sequential chains - Build reliable orchestration on top of Temporal — retries, timeouts, partial failure recovery, idempotency - Structured output enforcement: JSON schema validation, retry loops on malformed LLM output, graceful degradation - Tool call design: schema design that LLMs actually follow reliably across providers Evaluation & Harness - Own the eval framework end to end: ground truth datasets, automated scoring pipelines, regression detection on every PR - LLM-as-judge pipelines for qualitative output assessment - Latency regression testing — p50/p95/p99 tracked across every deployment - Adversarial test case design: ambiguous queries, missing data, conflicting sources, malformed tool responses Infrastructure - Model serving and cold start optimization - Async worker architecture for parallel sub-agent execution - Observability: trace every token, every tool call, every synthesis step Why This Role The problems here don't have blog posts about them yet. Parallel agent DAG execution under hard latency budgets, streaming synthesis across partial sub-agent results, eval harnesses for non-deterministic multi-step systems — these are genuinely unsolved at production quality. Small team. High ownership. Every engineer's decisions ship to production. Who We Want to Hear From You've shipped inference systems at: - A real-time AI product (search, coding assistant, chat at scale) - A model serving infrastructure company - An agent platform (any domain) Or you've built eval/harness infrastructure that a team of 10+ engineers actually trusted to catch regressions. Apply Send to: [careers@arcana.io] Include: One system you built where latency was the primary constraint — what you measured, what you changed, what moved Link to anything public (code, writing, talks) — optional but useful No cover letter required We respond to every application.

Requirements

You've built something that runs in production at a meaningful scale and you understand why it's fast (or why it isn't).
You've worked on inference pipelines where TTFT was the primary metric and you moved it meaningfully
You've built multi-step agent systems and you know where they break — not from reading papers but from watching them fail in production
You've written eval harnesses from scratch and you have opinions about what makes a ground truth dataset actually useful
You've debugged LLM non-determinism in production and built systems resilient to it
You've worked with streaming LLM responses and built infrastructure around partial output handling

Nice To Haves

You've fine-tuned models but haven't shipped inference systems
You've used LangChain/LlamaIndex but haven't built the layer underneath
Strong ML research background without systems exposure
Stack familiarity (we care more about depth than match): Go, Python, Temporal, Kafka, PostgreSQL, Docker

Responsibilities

Drive TTFT below 400ms for multi-step agent pipelines
Streaming optimization — first token to user while sub-agents are still running
KV cache strategy, prompt compression, dynamic context window management
Multi-provider routing: model selection by latency, cost, and task type across OpenAI, Anthropic, Gemini, and open-weight models
Design and implement Plan-Execute-Synthesize pipelines that run sub-agents in parallel DAGs, not sequential chains
Build reliable orchestration on top of Temporal — retries, timeouts, partial failure recovery, idempotency
Structured output enforcement: JSON schema validation, retry loops on malformed LLM output, graceful degradation
Tool call design: schema design that LLMs actually follow reliably across providers
Own the eval framework end to end: ground truth datasets, automated scoring pipelines, regression detection on every PR
LLM-as-judge pipelines for qualitative output assessment
Latency regression testing — p50/p95/p99 tracked across every deployment
Adversarial test case design: ambiguous queries, missing data, conflicting sources, malformed tool responses
Model serving and cold start optimization
Async worker architecture for parallel sub-agent execution
Observability: trace every token, every tool call, every synthesis step