Executive by trade. Engineer by training. Inventor by habit.
I’ve owned $1B+ product portfolios and have hands-on experience with AI going back to 2008. Through Aras Peak Group, I join your leadership team at the intensity you need, priced on outcomes.
Outcomes, not hours.
Priced on the results we agree on. A partner, not a vendor.
The right answer, even if it isn’t me.
If someone else is the better fit, I’ll say so, and help you find them.
Owned the number.
$1B+ product portfolios, mergers and acquisitions, multinational teams.
In the room where it’s decided.
Board, executive team or architecture review: I sit with the people who make your business run, and make their calls sharper.
Chosen, not chased.
I take on a few engagements a year, the ones where I can move the number.
Outcomes, with the years they took.
Computer-vision thesis, a co-founded research lab, and a paper at ICPR.
MASc, Electrical Engineering, University of Victoria (Applied Computer Vision Algorithms lab). Before that, degrees in electrical and nuclear engineering.
Built a SCADA-based lighting-control interface that became a core product.
Douglas Lighting Controls, later acquired by Panasonic. Joined as an application engineer.
Patented occupancy sensors that learn which sounds mean people are present, then launched the sensor line.
Douglas Lighting Controls. Filed 2011, granted 2014. Then led the launch of a patented occupancy-sensor line.
OEM and ODM partnerships for the hardware line: +5 points of gross margin.
Video infrastructure at Avigilon, part of Motorola Solutions from 2018.
Video infrastructure grew 20%+ a year, faster than any other business unit.
Avigilon, Motorola Solutions.
Led the industry's first AI NVR: +30% incremental revenue.
A network video recorder with the AI built in. Director of Product Management, AI Platforms, Avigilon.
Top Invention of the Year 2021 at Motorola Solutions, for the patent on access control that learns to predict intent to enter.
Avigilon, Motorola Solutions. Filed December 2020, granted October 2021.
Turned four brands' video roadmaps into one.
Johnson Controls: Exacqvision, American Dynamics, Illustra and Tyco AI.
Grew revenue at twice the market rate, every year.
Video and access control, Avigilon (Motorola Solutions). Recruited back by senior leadership in 2022.
Portfolio revenue +24% in a year; gross margin +3 points from pruning the line.
Revenue growth came from integrations with government and public-safety systems.
Product lead on several acquisitions, from diligence to fold-in and team integration.
Motorola Solutions.
Led product for a 300-person development organization, with 26 direct reports.
Senior Director of Product Management, video and access control. Avigilon, Motorola Solutions.
Recurring revenue from the installed base reached 20% of software revenue.
Simpler licensing the same year: net-new customers +30% in the first half.
Conceived and led the industry's first VLM-based Visual Alerts.
Vision-language models watching live video for situations described in plain language. Avigilon, Motorola Solutions.
Founded Aras Peak Group: fractional executive leadership, and enterprise-grade engineering from silicon to cloud.
Independent since October 2025. The firm offers fractional executive and board advisory, and builds its products with Conductor, an AI software delivery lifecycle with expert human judgment at the gates.
The thread through 18 years: computer-vision research, then patents, then shipped products, at the edge and in the cloud.
Six patent families, four peer-reviewed papers and three theses.
Network of dual technology occupancy sensors and associated lighting control method
On-device audio classification · edge AI
Networked occupancy sensors that learn, on the sensor itself, to classify which sounds mean people are present, using motion as the teacher.
Universal Lighting Technologies, Inc. · US8847750B1
Occupancy detection system and method having automatic adaptation capabilities
Adaptive learning from sensor data
Adaptive occupancy detection that learns a space's usage pattern from sensor history, then tunes its own lights-off delay instead of a fixed timeout.
Universal Lighting Technologies, Inc. · US8909382B1
Lighting control system, expansion pack, and method of use
An add-on expansion pack for a lighting controller, with set junction-box spacing, so larger lighting-control installations are simpler to plan and install.
Universal Lighting Technologies, Inc. · US10045451B1
Autonomous self-learning artificial intelligence intent system for access control
Machine learning · intent prediction
Top Invention of the Year 2021, Motorola Solutions
Access control with a machine-learning model that predicts when someone intends to go through a door, and labels its own training data from what happened.
Motorola Solutions, Inc. · US11157734B1
Clustering and active learning for teach-by-example
Active learning · video analytics
Clusters a video system's object detections by what they likely show, so labeling a few examples teaches the classifier about the whole group.
Motorola Solutions, Inc. · US20220301403A1
Method of generating a trigger initiating an emergency incident response workflow
When an emergency is reported, it gathers video and sensor data from that time and place, then starts the response workflow if similar conditions recur.
Motorola Solutions, Inc. · US12634398B2
A wearable ultrasonic step-distance meter with spoken feedback for blind pedestrians
A device worn on the shoe measures the distance between the feet with ultrasound at every step, adds up the distance travelled and reads it aloud to a blind walker.
Azad University (BASc, Electrical Engineering) · English title; the thesis is not in English
Fusing anatomical and functional medical images: CT and MRI with PET and SPECT, reconstructed in 3D
Software that overlays function from PET or SPECT on the sharper anatomy of CT or MRI and reconstructs the result in 3D, so it can be sliced and viewed from any angle.
Shahid Beheshti University (MSc, Nuclear Engineering) · English title; the thesis is not in English
Interdisciplinary Project-Based Learning in Ergonomics for Software Engineers: A Case Study
An ergonomics course project in which software-engineering students found where older users made errors with a software tool, then redesigned it.
IEEE · 3rd International Conference on Software Engineering Advances (ICSEA 2008) · 10.1109/ICSEA.2008.70
EasyDSP: Problem-Based Learning in Digital Signal Processing
EasyDSP, a custom framework that let university signal-processing students pose and solve their own multimedia problems alongside lectures.
Canadian Engineering Education Association (CEEA 2008) · 10.24908/pceea.v0i0.3751
Computer vision
A bed sheet with 144 pressure sensors whose readings are treated as images, so computer-vision techniques can detect breathing, apnea and movement at home.
University of Victoria (MASc, Electrical Engineering) · hdl:1828/1297
Work in progress: Problem-based learning in digital signal processing
Early results from adding student-defined, problem-based projects to a lecture-based signal-processing course, so the math meets real problems.
IEEE · 39th Frontiers in Education Conference (FIE 2009) · 10.1109/FIE.2009.5350718
Pattern recognition · computer vision
Turns bed-pressure readings into images, then uses computer vision and pattern recognition to tell normal breathing, sleep apnea and body movement apart.
IEEE · 20th International Conference on Pattern Recognition (ICPR 2010) · 10.1109/ICPR.2010.974
I’m a product executive: fifteen years in enterprise product, most recently at two Fortune 500 companies. I left with a question I wanted to answer with my own hands: if you started a software development process from scratch today, with AI, what would it look like? I spent six months building one. This is what I came up with, and the products it was tested on.
What I came up with
Fifteen years of enterprise product judgment, most recently at two Fortune 500 companies, built into an AI software delivery lifecycle.
Anyone can point AI agents at a codebase. The difference is when a person decides. Conductor, the process I arrived at, works like an engineering team that knows when to escalate. Routine implementation runs on a fast, lower-cost model; planning, blind test-writing and review run on the most capable one. When an implementation fails its checks twice, the final attempt moves up to the stronger model. When the builder meets a question it shouldn’t settle alone, it stops and asks. And every decision that is hard to reverse, such as a database schema, an authentication design, pricing or a deploy target, waits for a senior human. Expert time goes where it changes the outcome, and delivery stays cost-effective, predictable and high quality.
Nothing advances on a model’s say-so. Gates are exit codes, tests are written blind by a model that never sees the implementation, and a reviewer’s approval is bound to the exact commit it read. Each one-way door is red-teamed by adversarial critics, and the work does not start until the decision is signed; other work keeps moving meanwhile. The questions asked at those doors come from shipping enterprise products at Motorola Solutions, Johnson Controls and Panasonic.
Every run leaves a receipt in a ledger: the models used, tokens, cost, time, and the commit that was reviewed and merged. Across 118 features shipped in three products, that ledger is how I know the process works.
The products were the test. The process is the result.
The brief
A person decides what to build and what done means.
Plan
Agents break the work into slices and plan each one.
One-way doors
Anything hard to reverse, such as a schema, a deploy target or a credential, is red-teamed, then waits for a signed decision.
Build
Agents write the code against the plan.
Test
Automated gates decide by exit code; blind tests come from a model that never saw the code.
Review
Approval is bound to the exact commit that was read.
Merge
Refused if anything changed after review.
The read
A person reads the ledger, not a transcript: models, cost, time and the commit that shipped.
The trade between cost, speed and quality is as old as software. What AI changes is where the levers are. Six months of ledgers, checked against fifteen years of late decisions, found them in three places, none of them obvious, and each one a judgment call.
A team of senior humans reaches the quality corner, at a cost and a pace few projects can afford. Agents on their own are cheap and fast, and can’t be trusted with the outcome. The method’s shape is the judgment: it gives up the least on all three. Each lever below protects one corner.
Design the graph, not the pipeline. The hand that checks the work gets the same brief as the hand that builds it, and never the build itself. Nobody grades their own work. That is what makes an agent’s output trustworthy, and trust is what lets it run unattended.
Every retry costs time; every escalation costs more. A loop that asks too early makes a person the bottleneck; one that runs too deep spends a night going in circles. The depth sits where iterating starts to cost more than asking. Finding it is judgment, not configuration.
Not every step deserves the most capable model, and the steps that do are not the obvious ones. Spend expensive judgment, human or machine, where it changes the outcome, and let the rest run cheap. Model routing is a budgeting decision, not an engineering one.
Where the hands separate, how deep the loops run, where the expensive judgment goes: that balance is what I bring to your table.
None of these is a minimum viable product. Each was picked for a different layer, silicon to cloud, then taken as far as the engineering would go and held to the standard an enterprise would demand: security, architecture, performance, observability, quality. A sprinkler doesn’t need signed over-the-air updates, a file system that never loses a day, or a timeline that streams over a weak Wi-Fi link. It has them, because the question was never the sprinkler. The question was whether one process could reach that depth, in months, at a cost a small company can carry. Ten projects later, the answer is below, and one of them, esp32-seed, is that standard handed to anyone who wants it.
Firmware, radios, sensors, and AI models small enough to run on the device.
Real-time AI on a $2 microcontroller: radar classification that flushes only for the cat.
Drop-in, set-and-forget sprinkler controller: per-zone soil-water modeling and weather forecasts, no subscription.
ESP32 firmware framework: signed updates, Wi-Fi, storage, time sync and web admin, built in.
Keeps an iPhone's personal hotspot from shutting off by sending it periodic requests.
Distributed, cloud-hybrid systems: devices in the field, heavy lifting in the cloud.
Cloud-hybrid timelapse SaaS for distributed organizations: edge workers capture, the cloud renders.
Desktop software that turns one long recording into separate, named songs, even when there's no gap between them.
Multi-tenant SaaS products with real users and automated releases.
League-night SaaS for pickleball: deterministic matchmaking, private preferences, live court boards.
Household SaaS, installable from the browser, that answers: how long has it been?
AI that people can check, not just trust.
AI compliance assistant: every citation verified, word for word, against approved sources.
Fractional leadership is billion-dollar judgment at the intensity your problem actually needs.
Engagements run through Aras Peak Group: my judgment and leadership, plus senior people from my network, many from the Fortune 500 companies I've worked for, brought in when the problem needs deeper engineering, finance or investor relationships.
A short, flat-fee engagement. You get my honest read, including whether you need me at all, plus a scope and a compensation model tied to the outcomes we agree on.
Board, executive team, or both. You set the intensity; my compensation follows the metrics we agreed on.
With your leadership and your best people: strategy, operations, product, engineering and go-to-market.
Straw Booking opens once K's Cal.com link is set (CAL_URL in the Worker config).