I own the top of a B2B lead funnel at an industrial-AI company, officially titled Software Engineer. Before that I co-founded two ed-tech ventures and consulted for 20+ enterprise and e-commerce clients. I build AI products end-to-end — most recently DhanDost, architected so the language model never computes a number the user sees.
I decide from customer signal over internal opinion, and I’m specific about what my data does and doesn’t support.
4-hour hackathon · solo · 3rd of 15+ teams
Paste a job description, get interviewed by someone on screen who reacts to what you say, and receive a debrief that quotes your own words back. Built in four hours, designed for the failure case, and honest about what wasn’t built.
An AI coach that connects real transaction data and helps Indian salaried professionals decide where the next rupee goes. Architected around one constraint: the LLM proposes strategy in words, deterministic code computes every rupee.
My buyer survey returned under 1% response. Triangulating behavioural data instead surfaced something bigger: demand was running 6x toward a different product than the one we were positioning.
Institutional buyers weren’t sceptical of the programme — they were sceptical of the demand. So I inverted the motion: generate demonstrable student demand first, then approach the buyer carrying it as evidence.
Coursework · Product Management & Agentic AI, Vishlesan i-HUB IIT Patna
A full product requirements document and working prototype for an AI study-group
matcher — problem statement, persona, user stories with acceptance criteria, user flow, technical requirements for MVP and scale, and launch KPIs. Built as a certification project, so the persona is constructed and the metrics are targets rather than results.
A/B test on a paid-traffic product page, CTA placement as the single variable — ~1,500 visitors per variant, conversion 9.8% → 14.2%. One variable, clean sample, shipped the winner.
Surveys capture what people say. GA4, heatmaps and search capture what they do. When my survey underperformed, the behavioural stack became the primary instrument.
In DhanDost I removed the language model from the arithmetic path entirely. That cost capability and eliminated a class of user harm structurally rather than probabilistically.
Website-sourced leads grew ~28% year-over-year in GA4 while I owned site CRO. That’s lead growth, not a conversion lift, and traffic grew over the same window.
In Interview Cracker, every AI call has a retry, a timeout and a badged fallback. It fired twice on the day, for a retired model and a token cap. Users never saw either.