Product Builder · AI & Data Products
Hidden friction is the signal. I turn it into something teams can ship.
From silent sync failures to AI answers that should never be sent, I turn customer evidence into requirements, prototypes, and experiments teams can build and measure.
Career arc: Data Engineer at Wipro (2021–2022), Artificial Intelligence Researcher at AI Hub Canada (2022–2023), then Technical Product Analyst and Senior Technical Product Analyst at Method CRM (2023–2026). From February through August 2026, I built independent product projects and prepared for my MBA in Applied AI in Business, which started in August 2026.
Professional product work · Method CRM for Intuit QuickBooks
Making sync failures visible and recoverable
Support themes, workflow interviews, and telemetry showed that a successful sync and a failed sync could look identical to customers. I defined the requirements for visible sync status, customer-friendly error messages, and self-service recovery. The team launched all three capabilities.
Measurement: Monthly sync-error support-ticket counts declined after launch. Because this was an observational before-and-after comparison without a holdout or stable exposure denominator, I describe the direction but do not publish a causal percentage.
Professional product work · Method CRM onboarding
Defining activation around a value-producing workflow
I defined activation as the share of new SMB accounts entering onboarding that completed a predefined value-producing workflow rather than merely completing setup. A controlled experiment produced a 37% relative lift versus control; the exact workflow event, sample size, and test window are confidential.
Why this number is published: Unlike the sync analysis, this result came from a controlled experiment with a defined eligible population and control.
Portfolio simulation · synthetic conversations · not a production claim
Before AI answers, it has to clear the gate
The useful job: Confidence Gate turns approved support sources into a reviewable draft, exposes the evidence behind it, and keeps the support agent in control of what reaches the customer.
The hard boundary: Never invent policy, pricing, refunds, legal terms, or account state; expose another customer’s data; hide uncertainty; bypass permissions; or report success when an action failed.
The release rule: Missing, stale, conflicting, or weak evidence stops the answer. Sensitive intent and failed downstream actions route to a person with the reason recorded.
What I would try to break before launch
Stale sources, citations that do not support the claim, cross-customer leakage, prompt injection, false-success messages, and emotionally wrong responses in distress or hardship scenarios.
My product thesis: AI can accelerate the artifact, but the PM’s durable advantage is interpreting what people mean. Prototypes can replace many explanatory requirements—not explicit non-goals, permissions, failure behavior, telemetry, rollout, or “must never” rules.
Portfolio simulation · fictionalized context · synthetic data
Trust-First Retention
A redacted PRD, friction map, metric contract, experiment design, and working cancellation prototype with explicit failure behavior and trust guardrails.
Independent builds · implementation status shown for each project
What is implemented—and what is not
- TripPin AI: Functional Next.js prototype with mock client-side extraction and local browser state; no production users or live AI service.
- Matchday Intelligence: End-to-end local prototype with seeded data; not publicly deployed and no real users.
- Signal to Roadmap: Functional local demo using synthetic sample signals and cached AI responses; no production deployment or users.
- NPS Intelligence: Functional local Streamlit demo using 250 synthetic responses or an uploaded CSV.
- Retention Cohort Engine: Functional local Streamlit demo using generated cohort data; no production users.
- A/B Test Analyzer: Concept specification only; the repository does not currently contain runnable application code.
- Smart Onboarding Analyzer: Functional local Streamlit prototype using simulated funnel scenarios; no production users.
- AI Search Relevance Tester: Functional local Streamlit prototype using sample relevance data; no production search traffic.
Feature explorations · independent, unaffiliated concepts · not observed outcomes
More product thinking
Self-directed explorations applying PM frameworks to real products: opportunity sizing, user insight, solution design, and success metrics.
Slack exploration · independent and unaffiliated
A speculative source-linked layer for team chat that surfaces action items, decisions, and unresolved questions without presenting a generated summary as the official record.
Measures: decision-recall time, action-item completion, and critical extraction errors.
Duolingo exploration · independent and unaffiliated
A speculative recovery path that uses a learner-selected reason for an interruption and helps the learner earn momentum back through lighter lessons.
Measures: post-break churn, seven-day re-engagement, and low-quality lesson completions.