About

What I do

I work with engineering and business teams at some of North America's largest financial institutions, and increasingly with insurers, healthcare companies, and retailers, on one question: how do you take an AI agent from a Python notebook to a system a regulator can audit?

Most of my time goes into building, not talking. In the last two years that has meant 50+ working agentic systems across roughly 15 industries: KYC investigation agents, letter-of-credit examination that went from hours to seconds, underwriting triage, payment exception handling, retail order recovery. The customers stay behind their walls. The patterns end up in the field guides.

The through-line is governed autonomy. Reasoning belongs to the model. Policy, audit, identity, and time belong outside it. Getting that boundary right is most of the job, and it is the part the demos usually skip.

How I work

Receipts over adjectives. When a bank asks "which model should run this," the answer is an eval, not an opinion: cost per case, accuracy against a gold set, tool-skip rates measured rather than assumed, temperature sensitivity checked. I built that harness before agent evals were a market conversation, because the question was already in the room. One finding from that work: a mid-tier model matched the premium model on a KYC task at roughly one thirteenth of the cost.

I am deliberately not tribal about the stack. Claude, Gemini, GPT, Cohere, and open models all run in systems I have shipped, sometimes three providers in one process. Same for frameworks: LangGraph, CrewAI, and plain code all have a place, governed by a control plane that owns the institution's obligations. Where data sovereignty matters, and in Canadian banking it matters a lot, that means locally hosted models and agents behind privacy boundaries, which is also why I built a Cohere connector for process orchestration.

And I build in public where I can: reference implementations of the open agent protocols (MCP, A2A, AG-UI, AP2) before they hit the mainstream, each with a field guide, each honest about what is simulated and what is real.

Why "letmereviewyourcode"

The name started as a provocation. The gap between "this demo works" and "this system is trustworthy" is almost always visible in the code, not the slide deck. Reading a system carefully is the fastest way to understand it, and reviewing my own builds in public, including the decisions I would make differently, is the discipline this site exists to enforce.

It is also an invitation. If you are building agentic systems and want a second pair of eyes from someone who does this for a living, my LinkedIn is open.

The record

Twelve-plus years in tech and financial services across wealth management, banking, and capital markets. Computer science degree. Before the agentic era I spent years in the Salesforce ecosystem, including co-running the only hands-on automation workshop at Dreamforce 2022, and I hold certifications across Salesforce and Pega. Along the way I have influenced well over $50M in enterprise deal outcomes, co-authored Camunda's A2A blog series, been quoted in its open banking coverage ("not every problem deserves an agent" is a rule I still stand by), earned a Gemini 3 hackathon listing for a multi-agent mortgage system, and picked up a Value Pioneer award at Camunda's annual kickoff.

Off the clock

The discipline that keeps a bank's agents auditable scales down surprisingly well, and I have a growing passion for democratizing it: small businesses, mid-sized companies, and non-profits, especially Canadian ones 🍁, deserve a fair version of the AI conversation. What their tools actually do, what a model should cost per case, and where their data really goes. This is a passion, not a practice, and nothing is for sale. I wrote down the questions worth asking, and if you are wrestling with them, reach out.

Beyond the job

I am curious past the edges of the day job: a Perplexity fellowship, hackathons, connector submissions, video pipelines built with Remotion, and an unreasonable amount of time spent watching how the frontier labs' coding agents converge on the same primitives. When agents started buying things, I gave one my credit card under a signed mandate to see what would happen. It's documented.

Off the screen: a golden retriever named Mishka who reviews nothing and approves everything, a golf habit that keeps me humble, trips to Argentina with my girlfriend where the asado alone justifies the flight, and a love of cricket that started on childhood pitches in India.