What we have built, and what was measured.

The situation. A US Army prize competition sought fire control and flight guidance implementations that could run inside the constraints of the platform rather than alongside it.

The constraint. A guidance loop has to produce a bounded worst-case latency, not a good average. That rules out the general-purpose compute path before any benchmark is run.

What we did. Implemented the guidance loop in programmable logic: seeker fusion, state estimation, proportional navigation and autopilot, with timing as a property of the design.

What was measured. Competitively selected against a national field. Award value $25,000. Government-witnessed demonstration scheduled for October 2026.

Status. Awarded. Demonstration pending.

The situation. Perception and control for small autonomous platforms, where the compute budget is set by the airframe.

The constraint. Everything had to fit one module: perception, sensor fusion, and the control interfaces, with no cloud and no companion computer.

What we did. Built the stack on an AMD Kria KV260, with perception in fabric and the ROS 2 and MAVLink surfaces on the processing system. Flown on fixed-wing, ground and legged platforms, and extended to a five-unit airborne swarm.

What was measured. Roughly ten watts for the full stack on the KV260. Preflight discipline verifies actual data flow rather than parameter configuration, which is the failure mode that grounds most integrations.

Status. Flight proven and in active use across programs.

The situation. We wanted our own throughput claim to rest on an instrumented program rather than on a vendor benchmark.

The constraint. Throughput claims are usually made on generation speed alone. Verification is where the cost actually sits, and it is the part nobody instruments.

What we did. Delivered a production platform with two engineers on an agentic pipeline, recording engineering hours against delivered source.

What was measured. 41,200 production lines in roughly 575 engineering hours, against a COCOMO II estimate measured in person-months. Certified and uncertified rates reported separately, because the certified band is the one enterprises purchase in.

Status. Delivered. Method published.

The situation. Running a useful language model where there is no cloud, no link, and a hard power budget.

The constraint. Weight memory traffic is what makes conventional inference expensive. Removing it means changing the numeric representation, not just the accelerator.

What we did. Built a ternary architecture with a constant recurrent state, so there is no growing attention cache, and wrapped it in a grounded retrieval harness with constrained decoding that cannot answer outside its sources.

What was measured. A 1.3 billion parameter model running on a single AMD Alveo U50, with four deployed knowledge bases across general, maritime, defense and industrial content.

Status. Deployed and running. Not a projection.