Aman Jaglan

I am a founding engineer at Arc Intelligence, steering the build of adaptive systems that learn in production, and I earned my Master's in Data Science from The George Washington University. My work sits at the intersection of reinforced continual learning and operational reliability—designing teacher-student architectures that reason like research partners and hold up under real incident pressure.

Research & Benchmarks

Each project below lives inside Arc's Compounded Intelligence program, where we treat field performance as the core scientific instrument.

• Teaching LLMs to Diagnose Production Incidents with ATLAS — first deployed ATLAS demo, aligning gpt-4.1 with a GEPA-driven teacher to surface true root causes in ITBench cascades and deliver structured propagation chains instead of symptom lists.
• ATLAS Online: Hyper-Efficient Online Optimization — layered GEPA atop our RL-trained teacher to reach a 165% lift in Pareto frontier scores within two hours and ~$10, proving reflective prompt evolution as an economical accelerator.
• Introducing ATLAS — open-sourced the adaptive teaching framework; documented +15.7% accuracy, +31% task completion, and 97% non-degradation across control tasks by coupling diagnostic probing with conditional instruction.
• The Era of the Outer Loop — articulated Compounded Intelligence: a verifier-driven outer loop that turns outcomes into transferable policies, defining metrics like LR, dROI, MTTL, and TTT for continual learning systems.
• τ²-bench Judgment Layer Results — advanced the Judgment Gap agenda with a 24.0% pass@1 on dual-control telecom tasks, beating GPT-4.1 by 33% and doubling o4-mini while maintaining reliability across user personas.
• Reinforced Continual Learning for an Agent-First World — demonstrated cross-domain teaching traces that drove a 54% completion rate on CRMArena-Pro policy cases and 69.2% precision on true violations, establishing RCL as the stack for organizational memory.