Hey, I'm Shiven. I'm a technical founder who figures out exactly what a system needs to achieve, designs the architecture, and then builds it. Most recently, I was AVP of Technology at Imarticus Learning from 2024 to 2025. I led a 25-person engineering team spanning six distinct functions: data, mobile, web, QA, product, and growth. It was a massive learning experience, scaling a platform used by 45,000 learners across more than 30 countries. Before that, I co-founded a company that was They wanted the team and the technology. Not the exact exit we originally drew on the whiteboard, but a genuinely good outcome.. Today, I run product and engineering at Generally Critical.
Building a demo is the easy part. Anyone can point an AI model at clean data and spin up something impressive by the afternoon. The real value lies in the unglamorous infrastructure underneath. It is the bridge into a closed, twenty-year-old accounting system, a scoring engine that delivers identical results for every single player, or a calling system that simply cannot misfire. That layer is unfashionable, takes a year to get right, and cannot be copied over a weekend. That is the work I do best, and it is what makes a product defensible.
I approach product strategy with the same rigour as technical architecture. I focus on what to build, what to reject, and when to hold back capital until retention numbers prove out. To me, product development and market strategy belong in a single seat. Growth should be wired directly into the software itself, meaning funnels are read straight from the code, and activation and retention are built as core features rather than dashboards bolted on as an afterthought. Everything mentioned here is shipped, live, and in active use. I have written an extensive breakdown for each project, detailing what it does and the complex problems nobody else wanted to touch.
What I'm running
- Generally Criticalco-founder · product & engineering
A daily benchmark for business judgment: one real scenario, three connected decisions, a score you can defend, ranked against everyone who took it. It has been live since May, and the players who find the loop come back about five times each. Scoring is fully deterministic, no model in the loop, because a ranking is only fair when everyone is measured identically. Every play sharpens a profile of how you decide across four dimensions, and that profile is the point: it compounds with use, and it is what the leagues, the reminders, and the hiring signal will be built on. Players never pay; the value is the judgment data underneath.
deterministic scoringper-player judgment model
Read the deep-dive→
Building next
- UserDriftin build
Product analytics that reads your codebase, not just your events. It maps the real user journey from the source, then tells you in one plain sentence which screen people quit on, with the fix drawn in your own app. It ties each drop-off back to the metrics that decide whether you grow, activation, retention, and CLTV, so a fix is ranked by the revenue it protects rather than the count of users who hit the bug. Every incumbent is blind to the code; that blindness is the opening.
codebase-aware analyticsprocess mining
Read the deep-dive→ - SeptHQin build
An AI marketing studio. Paste your website, seven specialist agents learn your brand and hand you a week of on-brand posts and ads, each with its reasoning attached and a human approving before anything ships. The agents aim at the numbers that decide whether marketing pays for itself, CAC, the MQL to HQL conversion, and CLTV, so the output is pointed at pipeline instead of vanity likes. The hard part was making them genuinely reason instead of returning confident constants: every judgement is schema-validated model output, so a score is a real score.
structured multi-agent pipelineschema-validated output
Read the deep-dive→
Earlier
- ZeroCrossLMS3 clients · 3,000+ learners
The white-label platform three external education companies run their whole operation on: content, payments, certificates, and open-ended AI grading of essays and branching simulations at production scale. Four portals, one multi-tenant engine, 30 programs live. These are real customers running their actual business on it, not pilots.
open-ended AI gradingmulti-model evaluation
Read the deep-dive→ - TallyKaroresolution in seconds, not a phone call
A bridge into Tally, the closed twenty-year-old accounting system nearly every Indian small business runs on. An Electron connector reads it over ODBC, syncs to the cloud every fifteen minutes so the numbers are there even when the owner's PC is off, and answers questions in Hindi over WhatsApp. What used to be a multi-step lookup in a desktop app, or a call to the accountant, is now one question answered in seconds. Getting the data out at all is the moat; everyone else builds on data they can't reach.
Tally ODBC + XML bridgeoffline-first sync
Read the deep-dive→ - OnTune AI1,000+ minutes of live AI calls
A voice-AI platform that makes and takes real phone calls, applied to hotel reservations. I owned the infrastructure: a LiveKit realtime pipeline, and an AWS-serverless orchestrator that uses Redis to guarantee a tenant can never dial past its concurrency limit. More than a thousand minutes of live calls have run through it, booking, answering, and resolving queries end to end without a human on the line. Every call is transcribed and scored for sentiment and intent, so a manager sees which conversations went well and which need a callback without replaying a single recording.
Redis-throttled dialerLiveKit realtimesentiment + intent scoring
Read the deep-dive→ - TrueCarbon Nexusverification in 30 days, not 18 months · AshTerra Sciences
A dMRV platform, digital measurement, reporting, and verification for carbon restoration. Satellite imagery, IoT sensors, and field data feed an AI pipeline that verifies a project in 30 days instead of the industry's 18 months, and every project carries a public integrity score with a hard floor that fails it on its worst dimension rather than letting a strong composite paper over a fatal weakness. Gate, don't average.
dMRV verification pipelineintegrity-scoring rubric
Read the deep-dive→ - Arogya Naribrowser-native 3D anatomy
A women's health platform that teaches through interactive 3D anatomy you can turn around and look inside, running in the browser with nothing to install, in English or Hindi. The hard part was making medical 3D survive a cheap phone and a weak signal, for exactly the person most health software ignores.
in-browser 3D anatomymultilingual course model
Read the deep-dive→ - TruWealthfinancial literacy, from a locked statement
A wealth advisor for India's mass affluent. You upload the password-protected fund statement your registrar emails you, it reconstructs your whole portfolio in memory, and a fourteen-factor model scores each holding hold, review, or exit. The point is teaching: it shows a first-time investor why each call was made, in plain language, so they learn how their own money works instead of being told what to do. It never keeps the decrypted file and never sells on its own.
in-memory CAS parsing14-factor fund-quality model
Read the deep-dive→
Writing
Field guide
Wired for Growth
A field guide to the growth frameworks and metrics that actually move the needle. Not out here yet.
Off the clock
I'm based in Bengaluru. I build in long, focused stretches, care about how things are made down to details most people would call excessive, and read constantly about how other people built the hard thing before me.