Research AI Engineer: Agentic Platform & Evaluation

Research AI Engineer:
Agentic Platform & Evaluation

AIDAChip is the Alignment OS for AI-native silicon engineering teams — the first AI platform built for chip design teams, not just individuals. Chip design today is fragmented across disciplines, tools, and handoffs; we replace that with a multiplayer AI layer that keeps intent, knowledge, and execution coherent from specification to tapeout.

This role owns two tightly coupled areas at once: building our agent fleet and proving it works. You'll drive both the platform engineering (the agent machinery) and the research science (evaluation and publication) — owning the machinery that turns a domain expert's agent design into a shipped production agent, while owning the rigorous testing that guarantees its performance.

• Build the agent factory (harness, packaging, tool wiring) so a new role agent ships in weeks, not months
• Design and ship orchestration agents (workflow coordination, verification orchestration, cross-discipline handoff)
• Hold the fleet efficiency bar: minimize turns/task and agent token overhead, keep critical paths sub-second
• Design the human-eval program with chip-design experts; build auto-evaluation that reaches parity with human judges
• Research agent consistency: variance across repeated runs, drift across model swaps
• Instrument the experience signals that gate the product: task completion, abandonment, HITL integrity
• Pair daily with chip-design domain experts: they define what each agent must do and how it's judged; you ship it and measure it

• Shipped multi-agent / LLM-agent systems to production: tool-use harnesses, orchestration, durable long-running tasks
• Deep cost/latency/reliability discipline for LLM systems at scale, with strong backend engineering
• Rigor in eval methodology: inter-rater reliability, paired statistical testing, LLM-judge calibration
• Production engineering ability: evaluation work ships code into CI gates, not notebooks

• Dealing with vendors and mass data collection systems
• A PhD (ML/NLP/IR or adjacent), an exceptional publication record, or an exceptional shipping-systems record

• Evidence over vibes: every behavior ships with a witness; claims verified before asserted; systems fail closed and loud
• Measured, not narrated: a score not in the eval database didn't happen; benchmarks gate merges
• Agent leverage: each engineer directs fleets of AI agents, shipping at a multiple of typing speed while holding the quality bar
• Human-in-the-loop by design: agents propose; humans approve what matters
• Greenfield ownership: a whole North-Star workstream, not a slice of a slice — on a production platform with real enterprise silicon customers, not a demo