Product Work on Arjun Uvacha

Shipping AI that gives support agents 40% of their time back

Wed, 01 Jul 2026 00:00:00 +0000

The problem

Enterprise helpdesks drown in repetition. Agents spend most of their day re-answering questions that have been answered before, and employees wait on tickets that never needed to exist. With 700+ client organisations varying wildly in size, industry, and HR maturity, “just add AI” wasn’t a strategy — the AI had to produce outputs agents could actually trust and send.

What I did

Shipped AI-powered creation workflows. I designed and shipped the AI Assistant — it auto-drafts contextual responses, generates structured diagnoses, and surfaces resolution suggestions — and the Super Agent, which resolves employee queries before they ever become tickets.

Built the knowledge layer that makes the AI trustworthy. I designed Helpdesk’s knowledge/context graph — 16 entities and 80 evaluation metrics — the structured context layer that grounds every AI-generated output. Generic LLM answers become context-aware, client-specific ones.

Earned the right to ship by listening first. I ran the module’s largest-ever client engagement: 11 city visits and 100+ product gaps surfaced, then used usage metrics and feedback loops to drive prioritisation. That research also powered a UX revamp to make a complex enterprise tool effortless for non-technical users — while managing conflicting priorities across 39 client organisations.

Pushed how we build, not just what. I ran an AI dev POC where structured, PM-authored PRDs directly generated functional end-to-end code — compressing the gap between product definition and working software.

Ran the pod. Led a cross-functional pod (engineering, design, QA, client success) to 100% on-time, zero-defect releases, balancing CPO/CTO direction against enterprise client needs and engineering velocity.

The impact

The AI workflows are validated to save 40% of human agent time — 2 in 5 support interactions eliminated entirely, at the scale of 50,000+ end users. Adoption is tracked and iterated continuously across 700+ clients.

What this says about how I work

I don’t treat AI as a feature checkbox. The 40% number exists because of the unglamorous parts — the knowledge graph, the 80 evaluation metrics, the 11 city visits — not despite them. I build the trust layer first, then the magic.

From a hunch to a greenlit product line — with a prototype I coded myself

Mon, 01 Jun 2026 00:00:00 +0000

The problem

Employee Relations — grievances, disciplinary processes, workplace investigations — was a gap nobody owned. Clients were stitching it together with spreadsheets and email. No mandate existed to fix it; I had to create one.

What I did

Found and framed the problem. I identified the problem space and conducted primary research across enterprise clients to understand how ER work actually happens today — and where it breaks.

Built the business case. Market analysis, competitive benchmarking, and sizing that put a number on the opportunity: $500K ARR potential.

Prototyped it myself. Instead of pitching with slides, I vibe-coded a validated, working prototype end-to-end using AI tools. Stakeholders didn’t have to imagine the product — they clicked through it.

Navigated the greenlight. Building conviction without authority: I took the case through CPO, sales, and finance approval to formally greenlight the product line. Engineering build is now in progress.

What this says about how I work

0→1 isn’t a phase of my career — it’s the muscle I keep using. And AI tools have changed what a PM can bring to a greenlight meeting: not a deck describing a product, but the product itself. That’s the kind of leverage I bring by default.

Growing weekly retention 6.5× by fixing the wait

Fri, 01 May 2026 00:00:00 +0000

The problem

Phable connected patients with doctors for chronic-care management. Patients kept churning — and the interviews told us why: the experience around waiting was broken. Patients had no visibility, doctors had no predictability, and every side of the marketplace blamed the other.

What I did

Started with the users, not the roadmap. I ran 40+ user interviews across both sides of the marketplace and used them to define the consumer product strategy.

Managed the marketplace tension. Doctors (supply), patients (demand), ops, and leadership each had competing priorities. I shipped a patient queue experience that gave patients visibility and doctors predictability — without asking either side to absorb the other’s pain.

Kept shipping on the money path too. In parallel, I built a payments product that cut Ops turnaround time from 66s to 28s, driven by data analytics and iterative delivery.

The impact

Weekly retention grew 6.5× — from 10% to 65% — a direct measure of engagement and delight at scale, in a category where users only return when the product actually earns it.

What this says about how I work

Retention problems are rarely feature problems. The queue experience won because the research pointed at the emotional core of the churn — uncertainty — and the product removed it for both sides of the marketplace at once.

0→1 with no engineers: a 10× program built from no-code parts

Mon, 20 Apr 2026 00:00:00 +0000

The problem

Edureka served working professionals — but ~1.5 million fresh engineering graduates come out of Indian colleges every year without industry-ready skills. A full-stack web development program for them was a big opportunity in uncharted territory: a segment with low ability to pay, high tech-savviness, and no history with us. The mandate: validate it fast, spend almost nothing.

What I did

Built the MVP from no-code parts — in 3 days. With no engineers and one support person, I assembled the entire acquisition product myself: an Unbounce landing page (heavy on FAQs so we didn’t drown our lone support rep), GoToWebinar demos, Zapier automations wiring registrations, attendance, and demo poll responses into Google Sheets, and a standalone Razorpay payment page. A three-step funnel — register, attend a demo, pay — filtered genuine candidates from browsers.

Let the first month teach us. ~47 enrollments at ₹10,000 each (~₹4.8L), at a CPL of ₹40 and CAC around ₹6,000. Enough signal to invest in iteration.

Optimised every stage of the funnel. I broke it into registered → attended → interested → enrolled and interviewed ~50 students across stages. Traffic-to-registration went 8% → 13% (mobile-first, asset-light page, first-fold CTAs, “job-ready” copy, participation certificates, multiple demo slots). Attendance went 30% → 42% (reminder emails and SMS at T-1 day and T-30 minutes). Payment conversion went 2% → 3.1% (automated lead scoring on background and demo engagement, drip campaigns, a ₹1,000 advance-payment page, EMI and buy-now-pay-later options).

Productionised only after the formula worked. Dedicated landing page, Zoho CRM journeys, marketing campaigns, per-user payment links — built with engineering once the hacky version had proven every assumption.

The impact

Leads grew from 6,000 to 18,000 a month; traffic-to-enrollment conversion went 0.05% → 0.17%; monthly revenue grew roughly 10× to ~₹50L by month six.

What this says about how I work

Validation doesn’t need a roadmap slot. The fastest way to answer “will this work?” is to duct-tape a real funnel together, put real money through it, and let each month’s data pick the next iteration.

The landing page that was selling the wrong thing

Fri, 10 Apr 2026 00:00:00 +0000

The problem

Edureka’s postgraduate programs (AI/ML with NIT Warangal, Data Science with IIT Guwahati) had a landing page problem that showed up everywhere: conversion below market peers and below other Edureka pages, sales reps fielding the same questions repeatedly, poor page speed, and little love for the design. As marketing spend scaled, the leaks got expensive.

What I did

Research before pixels. Qualitative: 15 interviews with customers who’d bought the program, 300+ sales call recordings across ~70 buyers at different funnel stages, and 7 interviews with sales and ops reps. Quantitative: Google Analytics and PageSpeed work on conversion by channel, bounce, time-on-page, and device split.

Found the real purchase driver. ~90% of buyers chose the program because of the university collaboration — which the page barely emphasised. Buyers were upskilling for promotions or career transitions, always comparison-shopped 3–4 providers, and needed credibility signals, not feature lists. We were feature-selling; they were value-buying.

Rebuilt the page around credibility. University-verified curriculum, testimonials, instructor profiles, career assistance services (mentoring, resume building, mock interviews), live instructor-led training, lifetime access, 24×7 support — and a proper FAQ section answering the questions sales kept hearing.

Validated before committing. A 50/50 A/B test for 30 days on the AI/ML program, tracking conversion, session time, bounce, scroll depth, and engagement. The new variant won decisively and shipped for good.

The impact

Traffic-to-lead conversion rose 3.6% → 4.8%, lead-to-payment 0.9% → 1.4%, and ‘hot-lead to sale’ turnaround dropped 33% — from 30 days to 20.

What this says about how I work

When a page underperforms, the temptation is to redesign it. The leverage is almost always in understanding why people actually buy — and 300 call recordings are a cheaper teacher than three redesigns.

Demo-to-enrollment: fixing the moment parents decide

Fri, 20 Mar 2026 00:00:00 +0000

The problem

Cuemath teaches math online to KG–8th grade students. Plenty of parents booked and attended demos — then didn’t enroll. My mandate: improve the demo-done → paid-enrollment conversion, the most expensive leak in the funnel.

What I did

Instrumented the funnel first. Broke it into demo experience → interest → enrollment and analysed conversion at each stage: attendance, engagement, interested-to-payment, time-to-interest, time-to-payment, demographics, devices, post-demo email open rates, even the share of demos with technical issues.

Then sat with 35 parents. Their questions, motivations, and barriers mapped to three recurring frictions. Customization: every child is different — shy or outspoken, Olympiad prep or fear of math — and the standard demo diagnosed none of it. Teacher credibility: trust in the teacher’s ability to teach, hand-hold, and stay competent was the deal-breaker. Operational disconnects: parents didn’t know a laptop was needed, sales counselling felt generic, and one parent attended while both decided.

Shipped in three phases. Phase 1, pre-demo personalization: captured the child’s problem areas, parent preferences, and urgency before the session, surfaced on a teacher dashboard so instructors could target the live demo. Phase 2, post-demo engagement: a parent dashboard with the child’s certificate and diagnostic report, math puzzles, the Cuemath method, and full teacher credentials — plus a teacher assessment form capturing the child’s skills, attention span, and persona alongside parent signals like fee sensitivity and urgency. Phase 3, lead scoring and routing: parent preferences and teacher remarks flowed into the admission counsellor platform, so counselling stopped being generic.

The impact

Demo-to-enrollment conversion rose 18% → 29%, and turnaround time dropped 30% (2.9 → 2.02 days). The teacher, parent, and counsellor were finally looking at the same picture of the same child.

What this says about how I work

Conversion problems at the bottom of a funnel are usually empathy problems in disguise. The fix wasn’t a better pitch — it was making sure nobody in the chain had to treat a specific child generically.

Two experiments, two honest verdicts

Tue, 10 Mar 2026 00:00:00 +0000

Experiment 1: the Brainly channel

Hypothesis: a meaningful share of math learners on Brainly would enroll with Cuemath. Setup: a Cuemath landing page on Brainly, a co-branded sign-up flow, exposed to 50% of Brainly’s North American math learners.

Observed: 15–20 leads a day — a nice-looking number hiding a ~60% junk rate, weak interest in the programs, and registration-to-demo and demo-to-payment conversion far below every other channel.

Verdict: not an effective channel without a heavy upfront qualification step. Shut it down. The rule it left behind: an acquisition channel that skips qualification doesn’t remove the cost — it relocates it to the sales team.

Experiment 2: the freemium portal

Hypothesis: giving parents a free self-learning product would shorten and improve conversion. Setup: a new sign-up page, a freemium K-8 learning system, and four weeks of organic blog traffic routed at it.

Observed: the acquisition metrics flattered — sign-up conversion 5.2% → 9.9%, login rates 19% → 60%, 300+ registrations. The engagement metrics told the truth: week-1 retention was 20%, and demo bookings were under 1% versus 30% in the regular flow.

Then 30+ parent conversations explained why: parents wanted live human instruction, not recorded videos; they had neither the time nor the confidence to guide the child through a platform; they saw it as practice, not learning. A parallel 300-parent pilot confirmed the preference for a custom tutor over one-size-fits-all software.

Verdict: the product wasn’t ready for self-serve, and the demand wasn’t there. Scrapped it — including the parts I’d championed.

What this says about how I work

Experiments only earn their cost if you’re willing to believe their answer. Both of these produced numbers that could have justified continuing; the discipline was in looking at the metrics that measured value, not activity.

Building a construction SaaS 0→1, for users who'd never used software

Fri, 20 Feb 2026 00:00:00 +0000

The problem

Construction runs on paper, phone calls, and trust. At Vishwa Samudra Engineering — a 10,000+ employee conglomerate — approvals took weeks, material pilferage hid inside the industry’s accepted ~1% “unaccounted expenses” (about ₹1 crore a month on ₹100 crore of bulk procurement), quality had no real-time feedback, and site accidents were rising as the company scaled. I knew none of this domain when I started.

What I did

Earned scope with the boring work first. Started part-time, mapping offline processes to an org-wide ERP for procurement, contracts, accounting, and finance. Approvals went from weeks to days; weekly issues dropped from 22 to 3; resolution time from 3 days to 1; ₹400+ crore of monthly purchase and work orders flowed through the system. That earned the mandate — and a team of 10 (2 APMs, 4 BAs, 2 analysts, run on OKRs) — to go after the sites.

Set a four-part product strategy. Bridge the site-to-head-office gap; build in-house for material, quality, and safety while integrating external tools for progress tracking; unify everything in one insights platform; keep ERP for accounting and admin only.

Built the site suite 0→1, for non-technical field users. TrackZ (material): weighbridge-verified tracking covering 40% of all material received across projects. KwartZ (quality): 100+ lab test types and site inspections digitised — 15,000+ lab tests and 4,000+ inspections reviewed, result-calculation time down from 30 minutes to 3, real-time alerts to the chief quality controller. Safety Eye (safety): QR-based issue reporting plus trainings, checklists, and drills — 2,100+ issues resolved, resolution down from 25 days to 5. BeatZ (security): patrolling, material movement, incident and SOS reporting. WingZ (shutters): utilization tracking for shuttering material. On top: PowerBI dashboards and a Command Control Center — an industry-first real-time intervention system giving head office live control over site issues.

The impact

$11.5M in quarterly operations digitised, incidents below 10 a quarter, and site fatalities down to zero. The honest footnote: site-team adoption was a constant grind, and external developers’ estimation habits tested every timeline — building the culture was harder than building the software.

What this says about how I work

No data, no domain knowledge, no digital culture — strategy still has to come from somewhere. Mine came from walking the sites, quantifying the leaks, and sequencing trust: fix the head office first, and the sites let you in.