Training Engineer — Research Replication

Infinity Artificial Intelligence Institute is developing systems to optimize AI agents for reproducing research tasks. The Training Engineer will build replication-focused agent harnesses, train agents using supervised and reinforcement-learning signals, strengthen anti-fabrication evaluation, and scale GPU-based infrastructure for parallel benchmark runs.

Responsibilities:

• Autoharness: design and own a replication-optimized agent loop: multi-session resumable runs with state on disk (Ralph-style iteration), budget/wall-clock governance, self-verification against per-task test contracts, automatic stuck-detection and recovery. Today's runs die in ways a purpose-built harness would survive (idle wedges, environment preflight failures, lost context across sessions)
• Training: move beyond prompting. Mine our run transcripts (hundreds of graded attempts with per-task pass/fail) for supervised and RL signal: replication is a near-ideal RLVR setting because every reward is an executed test, not a preference label. Fine-tune/RL-train agents against the graph-completion objective
• Honest-by-construction evaluation: extend the behavioral anti-fabrication stack (isolated re-execution of the agent's reproduce.sh, wall-time/GPU-utilization checks, transcript forensics). Text-pattern safeguards get routed around; you'll build ones that require the agent to change its run, not its wording
• Scale the infra: parallel multi-paper runs on GPU nodes, watchdogs, cost telemetry, deterministic re-grading