AI research and tools for production systems

Kalmantic runs as an AI-Native organization powered by NemoClaw. We research how autonomous agents with persistent memory and scheduled heartbeats operate inside real enterprises. The brain is commodity. The harness accumulates.

Inference is constrained less by compute and more by memory movement, batching behavior, and context growth. We build tools (PeakInfer, PeakWeights) and publish research on why AI systems fail when usage grows and costs behave strangely.

SOUL.md, MEMORY.md, HEARTBEAT.md, AGENTS.md. Four markdown files define an agent. The innovation is persistence, not neural breakthroughs. Every day an agent runs, its harness gets richer. That is the moat. We research what makes harnesses compound and what makes them brittle.

LegacyCodeBench, AutoBench, TravelBench. Three active benchmarks across legacy code, automotive, and travel. We measure what production AI systems actually get wrong, not what they get right on synthetic evals.

Inference performance is a systems problem. GPU utilization, memory bandwidth, batch scheduling, network topology, and cost-per-token across deployment configurations. We research the infrastructure layer that determines whether a serving setup is viable at scale.