I learned how software breaks before I learned how to plan it.
Nine years turning messy, half-built products into shipped ones. I started as an engineer, became a PM, and now I build the agentic tools myself. Ideas to production.
You're looking at one of them. Scroll.
01 · QUALITY KIOSK · 2016–2020
batch-deploy failures across 700 ATMs, five countries
“My instinct in a mess was to build the system that made a whole class of errors impossible. I still work that way.”
Solution Engineer - 700 ATMs. 5 countries. One grid.
I deployed patches and fixes across 700+ ATMs in the UAE, Bahrain, Kuwait, Egypt, and Qatar for Mashreq Bank. Every machine was a different mix of vendor, model, variant, and software version, and each patch only fit certain combinations. Batch deployments were failing 20% of the time. So I mapped it. Vendors and models down the side, patch names across the top, a separate deployment page for every valid combination. The error rate dropped from 20% to under 2%. Each machine dispensed around 70,000 dirhams a day, so the downtime I killed was real money. Mashreq gave me a Certificate of Recognition for it.
Performance Engineer - Reading a system by its bottlenecks.
Then I moved to performance engineering, profiling client and server response times and working with developers to fix them. Page load and API times improved 10 to 20%. I learned to see a product by where it slows down, not by its feature list.
Senior Test Engineer - Owning quality where mistakes are expensive.
Then I led a 7-person QA team on AIA Philam Life's core insurance platform. 150+ test cases per BRD, 10+ critical defects caught per cycle, in an environment where a bug is a regulatory problem, not a ticket.
Four years in QA taught me the thing that redirected my career: most defects don't come from bad code, they come from bad requirements. So I moved upstream, to where the problem gets defined in the first place.
02 · SPINACH + DESIGNCOZ · 2021–2023
client engagements, banking to gaming
“I moved into product through design studios. Two years of range: shipping other people's hardest problems, across more domains than most PMs see in a decade.”
Product Manager - An Enterprise grade design system shared between two countries.
At Spinach Experience Studios I built and shipped a shared design system for a US insurance firm, spanning three teams across two countries, with full localization for the Brazilian market. I ran the discovery workshops, wrote the PRDs and user flows, and led a cross-functional team of eight designers to turn strategy into something teams could actually build.
Senior Product Manager - 50+ clients. 8 designers to 35.
At Designcoz I led product across 50+ client engagements, spanning banking, fintech, healthtech, edtech, ML, web3, and gaming. I stood up a conversion optimization service as a new studio offering, and both pilot clients stayed on for eight months and more, which proved it as a repeatable revenue line. I grew the team from 8 to 35+ designers and built the lead structure and performance management under it.
“Range is a skill. I learned to walk into an unfamiliar domain and find the shape of the problem fast.”
After that much breadth I wanted the opposite. One product, owned all the way through, where the hard part was the intelligence itself. That took me into health.
03 · FLUENT HEALTH · 2023–2025
user pilot for the AI health agent
“This is where AI stopped being a tool I used and became the product I built.”
Senior Production Manager - Founding PM. A beta way to test
I joined an early-stage healthtech as its founding product hire. The platform had been stuck in beta for a long time with no clear path to delivery. I brought roadmap structure to it and ran 200+ patient interviews and 30+ doctor sessions to ground every decision in a validated need. In about seven months we went from open-ended exploration to focused delivery.
The pivot: marketplace to health agent.
The validated need did not point at the marketplace we were building. So I made the call to pivot to an AI health agent. We built it on the earlier versions LangChain and n8n, integrated it with patient EMRs over FHIR R4 for structured clinical data, and shipped conversational health guidance with humans in the loop. We launched a 150-user pilot to validate the core use cases.
“I reached for AI because it solved a real patient problem, not because it was on a trend deck.”
Once I had built an agent people actually used, I could not go back to writing specs and waiting on a handoff. I wanted AI in the build itself.
04 · AEREM · 2025–2026
weeks per frontend sprint, cut from four. No added headcount
“Here I turned AI on my own team, and it changed the economics of shipping.”
Senior Product Manager - Two B2B platforms, owned all the way through.
I led product for two B2B SaaS platforms in solar: a partner portal for EPCs and installers, and a solar IoT monitoring platform. I managed two PMs, and together we shipped 40+ features, including Design Services as a new revenue stream and WhatsApp notifications as a paid add-on.
Analytics from zero. 250+ events.
I built product analytics from scratch across the partner and storefront products, Mixpanel and MoEngage, defining 250+ events, funnels, and dashboards. Once the team could see the funnels, I used them to lift onboarding and core flow conversion by around 20%.
Four weeks to one.
Then I brought AI-assisted development into the workflow. For select flows, frontend sprint time dropped from around four weeks to around one, with no added headcount.
“A PM who builds with AI does not just ship faster. They change what a team of the same size can do.”
By then the tools had caught up to the ambition. So I left to build with them full-time.
05 · INDEPENDENT · 2026–present
“Now I build and ship agentic products on my own. Not prototypes. Things running in production.”
Live pilot
Agentic Property Management System
A full stack PMS with an agentic concierge layer for real-time guest communication and AI-driven operational insight over a hotel's own data. Built on LangChain, LangGraph, pgvector, and Supabase. In active pilot with a hotel chain in Sri Lanka.
Open source
Chimes
A CLI notification tool for AI coding agents like Claude Code and Cursor. It plays distinct audio cues for task completion, pending questions, and token exhaustion, so developers stay in flow instead of babysitting a terminal. Public on GitHub.
Private
Price Intelligence with a self-healing agent
A scraping pipeline for pricing across quick commerce platforms, paired with a maintenance agent on the Claude Agent SDK. When a run fails, the agent diagnoses and repairs it on invocation, so there is no manual work in the data pipeline.
Public
Product analytics Skill
A custom AI Skill for product analytics work. It generates Mixpanel event schemas from template sheets, writes the tracking code, and reads a codebase to surface recommendations and annotations, with human-in-the-loop approval gates before anything hits production.
“I don't pitch agentic systems in a doc. I run them in production and watch what breaks.”
None of this works without a way of working. Here is how I actually build.
06 · HOW I BUILD
Every product I've shipped ran through the same loop. AI didn't replace the thinking. It collapsed the one step that used to cost weeks.
- 01
Define the problem, backed by data.
Every build starts with a problem statement, and it has to be grounded in data, never a hunch. Who is facing this? Where in the flow does it hit them? What do the numbers say? What is it costing the business? A drop in a funnel and a line from a user interview both count, as long as I can back them.
- 02
Open the floor for solutions.
I bring the problem to the team and we generate ways to solve it. Everything is welcome, from a whole new flow or an AI agent down to widening the click area on a button. The problem is fixed. The solution is still open.
- 03
Score and prioritize.
Not every idea earns a build. I score them with RICE or ICE, reach, impact, confidence, and effort, so the call is explicit instead of driven by whoever argued hardest. One thing shifts the math: when prototyping is cheap, the effort score drops, so ideas that used to look too expensive to try are suddenly worth a shot.
- 04
Prototype, then break it on purpose.
This is where AI changes the math. What used to take two weeks now takes a day, sometimes hours. I build the top-ranked ideas and put them in front of stakeholders and peers, and I ask them to poke holes. Better it breaks here than in production. If the prototype proves the idea wrong, that is a win, not a loss. I kill it here and move to the next one.
“AI didn't change how I think about product. It changed how fast I can test whether I'm right.”
- 05
Decide if it's worth building.
This is the call that matters most. With real feedback in hand, is this solution worth what it costs to ship? Sometimes the honest answer is no, and the best decision is to stop. If yes, we commit, and I define the events that will measure the success metrics I set, so nothing ships blind.
- 06
Make it production ready.
I work with engineers to take the idea from a working prototype to something that holds up under real use.
- 07
Ship and measure.
It goes live and I measure against the events I defined. Did it move the metric? If yes, on to the next problem. If no, I roll it back or rework it. Shipping is not the finish line, the result is.
MY SETUP
The loop is the same every time. This is the kit that runs it. Everything orbits the model.
One model per job, tuned for tokens
I don't run one model for everything. I frame problems with Claude Sonnet, where the reasoning carries the most weight. I code and build with Haiku or Claude Fable 5, fast and cheap enough to iterate hard. Testing runs on Google Antigravity and its own models, so Claude stays on building while something else tries to break it. The splits are mostly about tokens. Even the analytics work reflects it: a skill writes my Mixpanel events into a template, but I wire them in by hand instead of through an MCP, because the MCP would burn tokens on every call for something that takes me a minute.
A Claude Code setup built on custom commands
Claude Code is the center of it, and I've built commands I reach for constantly. mermaid-diagrams turns a flow into a flowchart. mixpanel-skill writes tracking events into my template. clean-code reviews a codebase and tells me what to optimize or cut. I skipped Cursor, one subscription I don't need, and run Google Antigravity alongside Claude instead.
I write the automation and keep context clean
I used to build automations in n8n. Now I'd rather write the code myself and let the model call it, the CodeAct pattern, which is easier to debug and not boxed in by what a visual tool supports. On long builds I clear context often with /clear so the agent isn't dragging stale history across pages, and I rebuild it with graphify when it needs the full picture again.
Human in the loop, heavily and on purpose
I'm hands-on by default. When a goal is clear, build a form with these fields, I let it run with /goals, and I use /loops to run tests and fixes on their own. The moment the goal gets ambiguous, I insert myself. With the analytics skill I read the full list of events it wants to create, make my tweaks, and approve before anything lands. The rule is simple: the more room there is to get it wrong, the less I let it run alone.
Why I built Chimes
I left an agent running, forgot to switch on auto mode while it was mid-edit, and walked away. I came back to find it had been stuck on step 2 the whole time. So I built Chimes. It plays an audio cue the moment an agent finishes or stalls, so I never lose an hour to a silent stall again.
What I don't hand to AI
I don't ideate with it. The first idea is mine, and then I use AI to build it out. And I don't let it make design decisions, because that is how you end up with something that looks like AI slop.
“AI writes the code. Taste is still my job.”
ASK ME ANYTHING
Ask me anything.
Type a question about my work, my decisions, or how I build. It answers live, grounded in my actual experience. Same skill I would bring to your team.
Chat with my agentCONTACT
Let's talk.
I'm looking for senior product roles on teams building with AI. If that's you, reach out.