Every post is grounded in shipped work: real architectures, real metrics, and honest failure modes.
Building a calculator where SIP/EPF/PPF/NPS and FIRE are first-class citizens — product choices from projections to inflation. Design, scope, and tests.
At Upcore I hunted 20+ prospects across 12 verticals; the output was a prioritized opportunity brief. The method behind turning interviews into a roadmap.
A handful of heartfelt, repeatable wins — checkout trust, search weights, PDP details — each small, each measured. The cadence beats the splash.
AI didn't remove the PM job — it moved the work up the judgment ladder. Specs get faster; decisions, scope, and risk reign. Real before/after inside.
The repeatable method I used across checkout, category pages, and forms: instrument first, form hypotheses second, never optimize an unmeasured funnel.
Wake word, speech-to-text, and a quantised Qwen 2.5 model in-browser via WebAssembly — no server, no API bill, no data left your device.
A 2,200-film dataset, gradient-boosted ML plus Bayesian priors, and a scoring engine that backtests itself. I built FilmRisk.AI — the failures included.
A Google Apps Script pipeline unifying Kibana, MongoDB, and GA4 into one auto-populated report — and what white-collar automation really costs.
A dormant 'Send Greetings' feature, an image-gen model, and a festival calendar across 27 occasions — feature engagement up 168%.
Why PRO+ users generated bills in Tally, not in-app — a two-week diagnostic that uncovered a 17:1 gap and changed the roadmap at C-suite level.
Website forms at 63.5% CVR but chat at 4.7%? We built a lead scoring router that sends gold leads to humans and bronze to automation.
Designing a notification system across renewals, transactional states, and conflict logic — with priority tiers and a daily cap users don't hate.
A 4-week sprint on GA4 custom events and Clarity heatmaps turned story-led category journeys into +17% PDP clicks, +34% leads, −16% bounce.
766 commits, 6 releases, one self-hosted macOS app. The honest design and failure modes of an AI recruiter agent — research and intros, not spam.
A field-by-field form teardown with GA4 events and Clarity — one field caused 23% of drop-offs. The fix raised lead submissions 124% in 28 days.
A 480K-session teardown: three separate checkout problems, three targeted fixes, two sprints — and a 26% relative drop in cart abandonment.