The ThesisTHE OPTIMIZATION IMPERATIVE · MARCH 2026
Every era of progress has the same shape: a new optimization target, a new data type, a better model of reality. Biology optimized survival on sensory experience. Culture ran on speech, then text, then the internet. LLMs consumed the written record of human knowledge, and that supply is close to spent.
EraOptimization targetData type
BiologicalSurvivalSensory experience
Oral cultureKnowledge transferSpeech
Written / printKnowledge scaleText
InternetKnowledge speedMultimedia
LLM eraLanguage predictionText corpora
World modelsPhysical predictionSensor streams of skilled action
The next target is physical prediction, and the ceiling on today's models is that they know language about the world, not the world. A plumber doesn't think in tokens when tightening a fitting. That knowledge lives in force, resistance, and consequence, and almost none of it has ever been recorded in a form a model can learn from.
So the binding constraint isn't architecture. It's structured, multimodal records of skilled action: motion, contact, intent, outcome. And collecting them is an operations problem before it is anything else. Consent, logistics, quality control, incentives, pay. I wrote the full argument in March 2026, then spent four months testing it with my own money. The work is below.
Everything below is the same bet made four times: find the physical-world data nobody is recording, build the cheapest honest way to capture it, and let the record compound into the moat.
Fluxor · real-world data collection for robot learning
Can demand from robotics labs fund the collection network that becomes the canonical dataset of skilled physical work?
One frame of a capture session as the pipeline sees it: hand pose on the frame; hand kinematics, task segments, and transcript-grounded contact events across the session below. Every layer from session FLX_20260331's real outputs.
Fluxor is the paper made operational. Every electrician bending conduit is generating exactly the manipulation data world models need, and none of it gets recorded. I built the collection loop end to end: onboarding tradespeople, bilingual homeowner consent, camera-kit logistics, SMS scheduling, upload, quality review, payouts. Behind it, a pipeline that fuses four IMUs with POV video and narration at 50 Hz, scores quality, segments tasks, weak-labels contact events, and exports to HDF5, LeRobot, RLDS, and Cosmos.
The capture model is tiered on purpose. IMU plus narration is cheap and physics-rich; a worker saying "tightening until I feel resistance" is an intent label at zero marginal cost; video layers in where demand justifies it. The bet is that dataset sales to robot-learning teams fund the network, with safety and quality analytics for contractors as the longer play. The clearest market signal so far came from the field: the CEO of one of the country's fastest-growing construction companies told me it's too early, and that what robotics teams want from collection changes every day. The tiered model is built to survive exactly that: keep the cheapest capture running until demand matures. And the pipeline is not the moat; anyone can sync sensors. What the network builds toward is the moat: a contact taxonomy — the structured vocabulary of how skilled hands meet tools and materials — and the data layer that robotics teams, safety products, and the trades' own tools all sit on top of. Every worker added makes it denser, one session licenses to many buyers, and analytics revenue grows linearly while the dataset compounds.
Where it stands: one fully instrumented session of real electrical work (my own), which became 24,657 processed samples at a 91.7 quality score with 227 candidate contact events. Contact labels, outside tradespeople, and scale are still unproven.
BarbellAI · sensor hardware and data systems for strength training
Can an instrumented coach put Olympic-level guidance in every gym, and become the physics ground-truth model for strength?
Top: the bench rig (two TI mmWave radars and a Jetson, mid-experiment) and the v6.1 clip board. Bottom: X-ray inspection from PCBWay's assembly line this week — the clip's radio module and pogo tail, and the full puck board.
Nobody remembers how strength is actually built. Load, bar speed, range of motion, the coaching call, and what happened next are never recorded together, so coaching runs on guesses. BarbellAI is the instrument to change that — a bar-mounted sensor clip and a rack-mounted mmWave radar puck, taken through v6.1 manufacturing. The contrarian call is no cameras: everyone reaches for vision to judge form, but a camera is blind in a crowded gym and makes the lifter perform for it — radar and inertial sensors see through the room and ask for nothing. The coaching product is the wedge — Olympic-level guidance for anyone, wherever they train. The compounding asset underneath is the physics ground-truth model for strength, built from millions of real reps no one else is capturing.
Validation runs like a lab. What would count as pass or fail gets written down before each capture, and the failed attempts stay in the record, because the failures are where the design came from.
Where it stands: centimeter-grade bar tracking proven on the bench rig, best clip-to-radar agreement r = 0.78 over 242 seconds, and a 294k-window radar training corpus assembled from seven public and simulated sources. Automatic load sensing and coaching efficacy are still unproven.
WarrantyBrain · warranty claims and operations, live on the Shopify App Store
Can every resolved claim make the next decision smarter, and can one operator run the whole company?
What I Found, the intelligence surface: every money-losing pattern in a store's claims, ranked by dollars, each with a reversible one-click action. Lead finding here is a $19,851 batch defect across 10 issues.
A merchant's warranty claims are their operations telling on themselves. WarrantyBrain ingests claims on Shopify, pulls the evidence (order, policy, photos, serial history), and recommends a resolution the merchant approves in one click, with refunds, replacements, and returns running behind reversible, audited guards. Claim automation is the wedge. The moat is what compounds behind it: a privacy-preserving fraud network that links signals across merchants without exposing any store's data, a decisions engine that turns that evidence into the right call, and a network effect where every resolved claim makes the next one — at every store — smarter.
It's also the proof of my operating model. I run product, engineering, support integrations, pricing, and paid acquisition alone, with the agent team described below.
Where it stands: live on the Shopify App Store, with money paths verified against a real store. Cross-merchant pattern intelligence is built; its network effect is still unproven.
I built booking, payments, waivers, and check-in for unstaffed strength stations in city parks (React, Supabase, Stripe), with family money committed, and took the pilot to Chicago Park District review. No precedent meant no process; the district never found the pilot a site. Shelved. The lesson stuck: when an idea has no category, standing up the approval path is part of the build, and it has to be scoped like one.
How I ThinkMETHOD
I start every problem by finding the standard answer, then asking why it's standard. Usually it's a playbook everyone inherited and no one re-derived — and that gap is where the work is. Six habits do the rest:
Question the standard.Strength coaching runs on guesses. Warranty claims get handled as one-offs. A tradesperson's skill dies with them. Each is a "that's just how it works" that doesn't survive a hard push.
Find the binding constraint.Not the loudest problem — the load-bearing one. Physical AI's isn't the model, it's the data. A dead ad funnel's wasn't the creative, it was a login wall. Fix the constraint, not the symptom.
See the pattern across domains.Four ventures, one shape: capture the data nobody records, let it compound into the moat. The domains look unrelated; the bet is identical.
Own the whole problem.I don't wait for a scoped task. I take the ambiguous, unowned version — recruiting through payouts, PCB through firmware, install through refund — and carry it until it works.
Prove it on the live system.Nothing is done until it's verified from the user's side. A second set of eyes with fresh context tries to refute every change before it ships, because speed without verification is how you ship confident nonsense.
Build two moves past the product.The wedge is what sells today; the company is the moat compounding behind it — the dataset, the taxonomy, the model. I build toward that from day one.
Run this way, one person operates like a team — and I think it's a preview of how every operations team will run. It all lives in private repositories, for commercial and IP reasons; I'll walk through any of it live.
Before This2016–2026
2025–2026J.P. Morgan, life-sciences banking. Portfolio execution across 120 companies, $100M+ in assets under management. 1 of 10 veterans selected nationwide for the SkillBridge Fellowship.
2023–2024U.S. Army recruiting command. Led 40 recruiters in Chicago; rebuilt the measurement and accountability cadence. Productivity rose 90% in a quarter.
2021–2023Field artillery officer, 82nd Airborne. Deployed to Poland during the Ukraine crisis; integrated artillery, air, and naval fires; 50+ live missions, zero safety incidents.
2016–2020West Point. B.S. in Business Management (Finance). Division I football.
