Execution
Ideas that arrive as working things
An invention isn't finished when it's clever. It's finished when it's in a specification, in silicon, and passing someone else's test. I've taken the same idea through all three.
AI evaluation · Production AI · Technical strategy · 3GPP · IEEE 802.11/802.15 · Bluetooth SIG · ETSI/FCC · San Francisco Bay Area, CA
Invent it. Prove it. Ship it. Then get the industry to agree on it.
I help technical teams turn difficult ideas into systems that can be measured, trusted, and shipped. My consulting work focuses on AI evaluation, production AI applications, and technical strategy — backed by Stanford and Caltech training in electrical engineering, 118 patents, hands-on product development, and years of building industry consensus across major standards organizations.
Wireless taught me that a technical claim is only useful when the test conditions are clear, the result is reproducible, and skeptical parties accept the measurement. I bring that same discipline to AI.
Where I work, in frequency
Log scale · 0.9–70 GHz
920 MHz sub-GHz
SUN OFDM and DECT NR+. Long-range, low-power links where propagation is generous but the band is crowded and the duty-cycle rules are strict.
2.4 GHz ISM
Bluetooth and BLE. Channel-access and listen-before-talk advocacy at the Bluetooth SIG.
5 GHz U-NII
IEEE 802.11 and narrowband coexistence. Priority-channel and channel-loading design; DFS radar detection.
6 GHz U-NII-5 → 8
Wi-Fi 6E/7 coexistence and enhanced dynamic spectrum sharing against an incumbent database.
6.5–8.5 GHz UWB
Secure ranging and localization. I lead the IEEE 802.15 work — 802.15.4z and 802.15.4ab — and the ETSI and FCC regulatory work behind it.
57–64 GHz mmWave
Radar coexistence, ranging, and direction finding at millimetre wavelengths.
Frequency
1
2
5
10
20
40
70 GHz
Now accepting pilot engagements
Frontier model standardization and evaluation
Frontier model evaluation has the same structure as radio conformance work, and most of the same failure modes: benchmarks that measure the wrong thing, results that don't reproduce across labs, thresholds set by whoever moved first, and no shared vocabulary for what was actually tested.
I have spent my career on exactly that problem — writing test conditions, defending pass/fail thresholds, and bringing competitors to a document they will all sign. I am now accepting pilot engagements in AI evaluation design, reproducibility, acceptance thresholds, and multi-party technical consensus.
What I take on
Six kinds of engagement
Wireless
Systems that survive contact with the real spectrum
PHY and MAC architecture from 920 MHz to 60 GHz — specified, simulated, and validated in silicon. Coexistence is the hard part, and it's the part I've spent my career on.
- Link budgets and coexistence analysis
- Ranging and localization accuracy
- Spec-to-silicon bring-up and validation
- Certification and interop strategy
Standards & spectrum
The rules your product will have to live under
Standards bodies decide what your radio is allowed to do years before you ship it. I write proposals, build coalitions, and take positions to regulators — and I can tell you where your influence is best spent.
- 3GPP, IEEE 802.11 and 802.15
- SUN OFDM, DECT NR+, 802.15.4z and 4ab
- Bluetooth SIG
- ETSI and FCC filings and responses
- Spectrum-sharing and access schemes
- Contribution and coalition strategy
Model evaluation
Measurement methods that hold up under scrutiny
Evaluation design for AI systems, approached the way conformance testing is approached in telecom: define the conditions, defend the threshold, and make the result reproducible by a skeptical third party. I build my own wireless channel models — body area, indoor, and outdoor — because the environment you test against is the evaluation.
- Channel models: body area, indoor, outdoor
- Benchmark and eval design review
- Reproducibility and test-condition specs
- Threshold and pass/fail argumentation
- Multi-party consensus process
Patents & IP
118 patents, and a working knowledge of what they're worth
Invention harvesting, claim review, prior-art reading, and portfolio strategy in wireless and signal processing. Available for expert analysis in standard-essential patent matters.
- Portfolio review and gap analysis
- Claim charts and prior-art assessment
- Essentiality and infringement reads
- Expert reports and analysis
Research
Hard problems, taken down to the mathematics
Stanford-trained in optimization, estimation, and statistical signal processing. I publish, and I hold results to out-of-sample evidence rather than to how good the story sounds.
- Estimation and Kalman/EKF sensor fusion
- Convex optimization and detection theory
- Algorithm design with latency budgets
- Peer-reviewed writing and tutorials
AI applications
I build the thing rather than describe it
I build AI applications end to end, from architecture and evaluation through deployment and operational handoff. Most are working prototypes built to answer a question faster than a deck could; one, a bilingual grounded-conversation app, runs in production today.
- React / Next.js / TypeScript / Postgres
- LLM integration, prompts, and guardrails
- Security review and hardening
- Test coverage and CI before launch
How I work
The part that isn't on a résumé
Depth
I go to the bottom of the problem
I'd rather spend a week on the underlying mathematics than a month on a workaround. When something is genuinely hard, that's usually the cheaper path.
Consensus
I like the room, and the room is the job
Standards are built in hallway conversations and revised text, not in slides. I enjoy the meetings, I read the opposing contribution properly, and I trade what I can afford to trade to get the thing passed.
Flexibility
Wide range, low friction
Sub-GHz to millimetre wave, PHY to product, committee to codebase. I'm comfortable being the newest person in a field and useful in it quickly.
The record
Verifiable
118Patents
13Peer-reviewed publications
6Standards & regulatory venues
StanfordPh.D. and M.S., electrical engineering
CaltechB.S., electrical engineering
Available for consulting, advisory, research, and expert engagements
Tell me what's not working.
The most useful first conversation is usually thirty minutes about the problem — the band you're stuck in, the standard that's about to bite you, the evaluation nobody trusts, the patent question, or the prototype that needs to become a product.