Houston, TX — Ph.D., Electrical & Computer Engineering

Jeffrey S. Jensen

I've built antenna swarms, drilling models and a self-driving pool system. Now I build AI agents, and the thing I care about most is whether you can trust them. Do they show their work, do they stay inside their permissions, and do they actually get better every loop?

01 — Research goals

Agents you can actually trust.

The idea is simple. If an agent is going to do real work for a real business, I want to know what it did, where it got it, and whether it was allowed to. That's agent reliability, evals and AI control, and it's where I spend my research time.

Reliability

Can the agent show where its answer came from? I audit research agents for the stuff that quietly goes wrong: claims with no source behind them, retrieval that missed half the picture, and gaps in the trace where nobody can tell what actually happened.

Evals

Freeze the evidence, change one thing, run it again. I set up research benches where agents loop through dozens of variations (tweak the MCP, tweak the guidance) and graders tell us if the output really got better. Then we check it against a fixed reference and run it again to see if it holds.

Control

Agents shouldn't be running loose. They read and write through MCP endpoints, they inherit the user's permissions and nothing more, the server enforces it, and anything risky goes back to a person to confirm.

Right now Running local models (Qwen, Ollama, OpenCode, MCP) through the same benches to see where they break: semantic errors, context recovery, and whether they do the same thing twice.

02 — What I've built

Sense, decide, act, measure.

  1. 2026 — Portage Founder We build agents for elite mid-market businesses. Think SEAL Team 6, not Accenture. Nightskye is our control plane: it deploys Claude and Codex agents with versioned bundles, run records, scheduling and a person watching. Our agents are doing real work today in mineral rights, tax-refund review and enterprise research.
  2. 2023 — 2026 Corva Chief Technology Officer Led ML for autonomous drilling in the Permian, from training base models all the way to closed-loop control on the rig. We grew from $20M to over $75M in revenue while I ran technology.
  3. 2019 — 2025 Clyr / Keto AI Founder Built a self-driving pool system: sensors in the water, software in the cloud, and the hardware on the pad that actually controls the pool. Three granted patents. Exited to private equity in 2025.
  4. 2016 — 2019 Arundo Analytics CTO · VP of Architecture Grew engineering from 5 to 70 people and built a big-data and ML platform for heavy industry.
  5. Earlier Siemens · Toshiba · Intel · Shipcom Industrial automation, systems architecture and leading engineering teams.
  6. B.S. · M.S. · Ph.D. Texas A&M University Electrical & Computer Engineering Computer-vision beamforming for swarms, funded by the U.S. Army Research Office. Built Medusa. 12 Under 12 Young Alumni, 2016.

03 — What I believe

What I believe.

About Clyr

  1. 01 A pool should basically drive itself. Sensors in the water, automation on the pad, and nobody babysitting the chemistry.
  2. 02 Remember these are pool guys. A tech should get through the whole visit with a few pictures and their voice. No forms, no hunting for buttons.
  3. 03 Get to the step before the step. Alerts are reactive. I want the leading indicator that tells you what's about to go wrong while you can still fix it.
  4. 04 Less is more. When the app turns into boxes in boxes, we pull it back to the agent and a clean screen.

About AI

  1. 01 If an agent can't show me its sources, I don't trust the answer.
  2. 02 Permissions live on the server, not in the prompt. An agent gets the user's access and nothing more.
  3. 03 Let the agents build autonomously. A person still says yes to anything that matters.
  4. 04 Build the eval before you believe the demo. Fixed reference, independent check, run it again.

About building

  1. 01 Close the loop. Antennas, drill bits, pool pumps, agents: it's always sense, decide, act, measure, repeat.
  2. 02 It has to work end to end. If it only works in the demo, it doesn't count.
  3. 03 Build it with off-the-shelf parts, put it in the real world and see what breaks. The simulation never matches the chamber.

04 — Medusa

Medusa in the anechoic chamber, Texas A&M, 2015.

Sixteen arms, one beam.

Medusa was the testbed for my Ph.D.: sixteen antennas on flexible arms, which is how it got the name. A camera with infrared depth sensing found every node and tracked it as the array changed shape, and the controller worked out the phase each antenna needed so all their signals added up in one direction. It ran at 2.48 GHz on off-the-shelf parts and boards I designed, and we measured it in an anechoic chamber.

The ripples on this page are sinc functions, sin(x)/x, which is the beam pattern of a uniform aperture. Nothing shows up until the wave gets there. Same idea as Medusa.

Medusa in the dissertation, p. 89 → Medusa 2.0: gr-medusa, HRG Lab →

05 — Ph.D. work

A Framework for Computer Vision Assisted Beamforming in Aperiodic Phased Arrays

Drone swarms and mobile radios end up scattered at random, and scattered antennas can't work together unless each one knows where it is. My idea was to use computer vision and depth sensing to find every node, track the array as it moves, and hand those positions to a controller that steers the beam. Then I built it and compared what we measured against the simulations.

Dissertation, Texas A&M University, December 2015. Committee chair: Gregory H. Huff. Supported by the U.S. Army Research Laboratory and Army Research Office, W911NF-09-1-0429.

Read the dissertation (PDF, 5 MB) →

Selected publications

Granted patents