P / 01Speed
Be fast or best, not both.
Pick a lane. Move quickly and learn in public, or commit to best-in-class at a slower cadence. Optimizing for both produces confused direction and mediocre output.
Full-stack product leader · A.D. 2026
Products that move markets. Teams that move mountains.
Currently thinking about
harness engineering for AI agents.
1 / 6
Open to senior product leadership & advisory
A decade building products across data, AI, and platform ecosystems — in environments where complexity is the constraint.
Execution keeps getting cheaper; what stays scarce is judgment in messy, high-leverage decisions. That's the work.
Built for environments where the systems are messy and the stakes are high.
I've spent a decade building and scaling products in environments where complexity is the constraint — multi-billion-dollar platforms, dense stakeholder maps, and heavily regulated domains.
The work has always been the same shape: turn data, AI, and platform strategy into durable business outcomes, not just shipped features.
- 0110+ years leading full-stack products across data, AI, and platform ecosystems in regulated environments.
- 02Directed portfolios underpinning tens of billions in annual revenue, with measurable gains in margin, speed, and competitiveness.
- 03Led cross-functional orgs of 100+ people across engineering, data science, design, and operations.
- 04Built and shipped 100+ prototypes across roles — from analytics engines to optimization platforms to experimentation systems.
- AI systems
- Platforms & economics
- Unit Economics
- Platform Strategy
- Pricing
- Underwriting
- Analytics
- APIs / SDKs
- PLG
- Neoclouds
- Frontier infrastructure
- Autonomous Innovation
- Digital Twins
- Physical AI
- Fusion
- Photonics
- Emerging Infrastructure
- Method & leadership
- Systems Thinking
- Product Strategy
- Experimentation
- Rapid Prototyping
- Outcome-Driven
- Psychological Safety
- Executive Advisory
Frameworks matter. Judgment matters more.
What separates product leaders in AI-shaped markets isn't more process — it's knowing what to build, when to say no, and where to concentrate scarce attention. The seven I return to most often.
P / 02Signal
Precision over accuracy.
A precise system — even if slightly off — can be tuned. An imprecise one is noise. Sharp goals and unambiguous criteria make strategy adjustments surgical instead of chaotic.
P / 03Experimentation
Discipline as capital allocation.
Experiments aren't a hobby. Timebox tests, set kill/ship thresholds upfront, watch the ratio, and treat consecutive losses as a signal to pivot — not a sunk-cost excuse to continue.
P / 04Problems
Painkillers before vitamins.
People reach for painkillers before supplements. Solve acute, expensive problems first; you earn permission to add delight later, and compounding trust along the way.
P / 05Systems
Systems over process.
A process is a checklist people follow. A system is an engine that produces outcomes even when you're not in the room. Invest in incentives and feedback loops, not more templates.
P / 06Bar
Reinvent only with a higher bar.
Rebuilding to arrive at the same place is waste. Reinvention is justified when it raises the bar by an order of magnitude — on reliability, speed, economics, or strategic control.
P / 07Outcomes
Outcomes over outputs.
Activity is not progress. The goal is maximum business and customer impact for the least necessary work. Focus teams on the delta you create in the world — behaviour changes, economics, resilience — not features shipped.
Interlude · on product sense
Walk as far as you can see — then decide and learn. Suddenly, you can see a bit further.
AI is no longer a feature to bolt on. It's a force reshaping how decisions get made.
My work sits at that intersection — designing harnesses, orchestration layers, and unit economics that make AI both reliable and viable.
AI as systems work
Designing systems that fail well.
Most AI agent projects don't fail because the model is weak — they fail because the system around the model isn't designed to handle failure.
Drawing on biomimicry, I focus on harness engineering: feed-forward and feedback guides, guardrails, and sensors that govern what happens when agents go wrong, so the system fails gracefully instead of catastrophically.
Harness around the agent
fig. 1
Architecture & economics
The moat is the orchestration layer.
The next advantage won't be "who has the best model?" — it'll be who owns the layer that routes across many models with the right economics and reliability.
That means treating AI infra like a portfolio: separating training capex from inference COGS, defining the true unit of value, and engineering routing, caching, and model choice so contribution margins improve as usage scales.
Model portfolio · routed
fig. 2
Frontier problems
Where infrastructure is breaking.
The real product opportunities lie where infrastructure is breaking: power and compute bottlenecks, memory and data exhaustion, networking and transmission limits, the missing safety layer for physical AI and trust rails for agentic commerce, and the architectural gaps in digital twins and fusion-scale systems.
I'm drawn to problems where solving the underlying architecture — rather than adding another wrapper — creates durable advantage for decades.
Bottleneck map
fig. 3
Every job is task plus judgment. AI speeds the task. Your judgment just got leverage.
Ideas in public, stress-tested at the edge.
A few recent pieces that best capture how I think — on LinkedIn, where most of the conversation actually happens.
№ 01
Product sense, in an AI-heavy world.
Three things: empathy for real users, understanding the system around the product, and naming the bet under ambiguity. Execution keeps getting cheaper — judgment in messy, high-leverage decisions is what stays scarce.
Read on LinkedIn
№ 02
Tackle the monkey first.
AI makes pedestals easy to build: prototypes, agents, demos. In Atlassian's State of Product 2027 survey, 80% of product professionals say AI helps them ship faster, yet customers see value no sooner. The monkey is the assumption the bet depends on. Test it first, in the smallest credible way.
Read on LinkedIn
№ 03
The bridge has to stand.
Whatever the requirements say, the bridge has to stand. Requirements are a lossy compression of judgment, and when an AI executes them, the spec is met and the outcome still fails. Intent-as-Source: one small, human-owned layer of intent, compiled into the requirements.
Read on LinkedIn
№ 04
When this fails, how does it fail?
On AI agents and harness engineering — why the real challenge isn't model quality, but designing systems that fail well. Inspired by biomimicry and benchmark gains from changing the harness, not the model.
Read on LinkedIn
№ 05
Nine billion tokens, split two ways.
A frontier model plans, decomposes and reviews; a quantized 9B model on a laptop GPU does the routine work. About a third of my nine billion tokens now run locally, at near-zero cost per token. Stanford's MinionS runs the same split, and NVIDIA's small-model paper argues for it.
Read on LinkedIn
№ 06
The spec economy.
An alumni talk — "Power Tools for Your Ambition" — on the next tool transition in product work. The premium moves to problem framing and decision quality; gunafood was the live example: define the product, make the calls, judge what ships.
Read on LinkedIn
Selected
highlights
highlights
- Granted patent — a method for accelerated material ageing. Experimentation isn't only digital.
- Advisory roles across multiple industrial companies on product, AI, and platform strategy.
- Internal recognition across employer programs for cross-functional impact and CSR leadership.
The lab.
GitHubJudgment compounds when it meets reality. A personal lab, run on my own time.
Open
product-skill
Thirty-plus product-management frameworks, packaged as an AI skill. PM judgment, encoded.
Live field guide
Open
intent-as-source
Intent as the source file, requirements as the build output.
Framework & paper
Open
andon
Code measurement an agent calls as it works. Every number carries its evidence and stays unconfirmed until an independent recompute agrees.
Recorded demo
Open
eigenorg
A research-grounded org simulator — 500 in-browser Monte Carlo runs that stress-test a structure before you staff it. Structure, not headcount.
Org simulator
Open
discern
A disciplined buying method an AI agent can run — independent-evidence clustering, per-claim confidence, counterevidence kept in view.
Method walkthrough
Open
laya-decision-assist
Decision assist on a 0.4B model, fine-tuned twice: one reviews product claims, one reads maintenance sensor flags, both on synthetic scenarios. Narrow models for narrow calls.
Model collection
Open
business-model-almanac
How a business runs, mapped: APQC's 13 process categories and 72 process groups, against 257 ranked software vendors.
Live atlas
Open
fintree
A 234-node, XBRL-mapped P&L ontology with a live explorer — the financial structure every FP&A team rebuilds, encoded once.
P&L explorer
Open
quantum-fusion-poc
VQE on tungsten–hydrogen binding, run on real IBM quantum hardware — a calibrated read on where quantum stands for hard chemistry.
Results write-up
The wider lab — skill runtimes, MCP servers, models, works in progress → github.com/gtm-k ↗ · huggingface.co/gtm-k ↗
Earning is a privilege; contributing back is a responsibility.
I focus my giving on expanding the circle of dignity and opportunity — from animal welfare and direct cash transfers, to children's futures and anti-corruption work that enables healthier economies.
- Humane World for Animalsadvocating for those who can't speak
- GiveDirectlycash transfers, agency intact
- UNICEFchildren, system failures
- Transparency Internationalcorruption as root cause
End of memo · let's talk
Building something ambitious? Let's compare notes.
Open to senior product leadership roles, advisory conversations, and opportunities at the intersection of AI, platforms, and complex systems. The best way to reach me is through LinkedIn.