How I actually work

Not frameworks collected from books — habits collected from escalation calls, courtrooms, and shipped software. Five principles and the operating stack underneath them.

Escalations are discovery, not damage control

The angriest customer on your worst day articulates the product gap better than any survey. I ran the Tier-3 escalation queue for a year and treated every threat-to-cancel as a research interview with unusually honest incentives. Three shipped patches and a company-wide protocol came out of that queue — and a 20% drop in P1 bugs reported by new clients.

Ship to the accent, not the demo

Dictation AI that works in a quiet office and fails on a Dogri-inflected sentence in a Srinagar courtroom is not a product — it's a demo with a rollout plan. Whatever the domain, I find its 'accent': the bad scan, the interrupted call, the clerk's workaround. That's the acceptance criterion.

The metric is a promise someone made

Behind '−35% time-to-resolution' is a clinic that got its problem fixed in 9 days instead of 14. I keep metrics honest by keeping the person attached to them. When a number improves and no human is better off, the number is lying.

Automate the loop, keep the human at the decision

In call intelligence, the AI listens, transcribes, classifies, drafts — and a front-desk human confirms. Removing that confirmation is a bigger product decision than adding the whole pipeline, and it should be earned with months of precision data, not shipped with confidence.

Anonymity over embellishment

The work on this site is anonymized — real problems, real mechanics, employers' products unnamed. I'd rather show you exactly how I think about a redacted thing than approximately how I think about a named one.

The AI-first operating stack

tools change monthly; the stages don’t

Discovery	Escalation queues & call recordings as primary sources · agentic deep-research for market scans · transcript mining with LLMs (patterns, not summaries)
Definition	One-page problem briefs · Miro for the messy middle · decision logs with discarded options kept visible
Design	Figma for the real thing · AI image models for lo-fi throwaways · clickable HTML prototypes over static mocks — if it can't be clicked, it can't be criticised properly
Delivery	Jira/Asana · acceptance criteria written as failure cases ('rejects a blurry scan gracefully') · weekly loops with support before anything ships
Evidence	Relative metrics tied to named humans · post-resolution CSAT · churn cohorts · 'would this survive the courtroom' as a standing review question