Ayush Kumar

PROBLEM
Male users didn’t receive responses from their sent invitations, so they didn’t experience the joy of dating, which led to drop-offs. This is reflected in key user metrics: low male retention, low order conversion, and low customer satisfaction (CSAT) scores.

APPROACH
We followed a design-thinking process and conducted in-depth qualitative user interviews. This unveiled an interesting finding: female users were serious, thoughtful, and deliberated before tapping the “accept” button. Our question was, how could we “casualize” their response behavior?

IMPLEMENTATION
“Casualizing” took us to the dating space where a user can say yes (swipe right) and no (swipe left) to a profile. We envisioned an MVP to be done only on the Android app (the maximum user engagement platform). We planned to use the existing APIs (using SOA). We integrated them into a new UI with a Tinder-like swipe functionality (implemented via a current component from the Android native library). I then sent it to production within two sprints (ten working days) and achieved ground-breaking success.

IMPACT
• Saw an 81% lift in responses.
• Increased order conversion by 5.3%.
• Jumped 4.8% in CSAT score.

PROBLEM
Repeat purchases constituted a small portion of eCommerce revenue, and we wanted to increase overall user retention and referrals.

APPROACH

1. Implemented a quantitative analysis of the cohorts of users buying products from us.
2. Obtainedualitative understanding of users—why/why not are they buying from us via in-depth interviews.

INSIGHTS

1. The primary reason people bought from us was the social component of our offerings, where each product sold was supporting a certain cause the users cared about.
2. We didn’t really “thank” them or made them feel proud about this across the whole user journey.

IMPLEMENTATION

1. Mapped the E2E user journey (pre-purchase, during, and post-purchase).
2. Began with the post-purchase user journey (thank you/order confirmation page and email).
3. Included inspiring content and videos to communicate the real-life stories of the end-beneficiaries and the impact they made just by purchasing from us.

IMPACT
• Saw a +3.5% increase in repeat purchases on our eCommerce shop.

KEY LEARNINGS

1. The hypothesis found by the human-centered design thinking approach was valid.
2. Replicate this “feeling of pride” across the other touchpoints of the user journey.

PROBLEM
We were only reaching a 35% connectivity with consumers through telecalls. This was caused because of the disposal of nonconnected calls and scheduling them for later.

APPROACH
We implemented an AI predictive dialer that will direct only connected calls to the team. Working with the data science and engineering team, I also created a requirements document along with all the use cases. We then developed an algorithm that learned based on various user demographics and activity parameters.

IMPLEMENTATION
Control Metrics:
• Ensure an agent is free before the call to ensure a good customer experience.
• A customer shouldn’t receive more than “x” number of calls in a day and should be called within the official time limits unless otherwise specified.
• Dialer will learn the answering patterns of users based on their demography and activity.
Other Work:
• Built reports and dashboards so that leads could monitor productivity and efficiency.
• Launched it as an A/B test with a few advisors and compared them against similar performing advisors in the other set.

IMPACT
• Increased telesales team productivity by 92%.
• Reduced the sales team size by half and added skills for cross-department usefulness.

PROBLEM

1. There were significant drop-offs across various stages of the donations flow.
2. Calculated the opportunity size to be a 15% increase (based on industry standards).

APPROACH

1. Identified the stages with maximum opportunity, and focussed on them.
2. Interviewed users in both buckets (successful and un-successful donations)
3. Followed design thinking approach of research, synthesis, ideation, convergence, prototyping, validation, execution & launch.
4. Identified the two largest drop-off points, and made the required UI changes.

IMPACT

1. The visitor-to-donation conversion increased by 8%.
2. The way forward was to replicate the same approach for the next two as well.

PROBLEM
The internet banking payment mode had a lower activation ratio. The primary reason was the bank’s site’s UX. The bank login page loaded in a web view on mobile devices, rather than having a mobile-friendly UI. This reflected in key user metrics: low activation ratio and order conversion.

APPROACH
We studied user behavior focusing on why users used net banking rather than credit/debit cards. We researched best practices followed for net banking by other merchants.

INSIGHTS

1. Net banking was primarily used by parents who perceived a security risk with credit card usage.
2. Best practices indicated the use of a vendor which converts a bank’s login pages into mobile-friendly pages
3. Also, a small user segment was ready to use UPI (unified payments interface) as an alternative mode of payment to net banking, which had a higher activation ratio.

IMPLEMENTATION

1. I spearheaded the integration of a new mode of payment: UPI with minimal handshakes and an intuitive UX.
2. I also led the integration of JUSPAY to make bank pages mobile-friendly and ensured database security.

IMPACT
• Increased activation ratio by 9.2%.
• Improved order conversion by 3.7%.

PROBLEM
While working with an online matchmaking platform, we wanted to diversify as a business into an online wedding planning platform. We were tasked to come up with a business model canvas.

APPROACH
It took a mix of primary and secondary research with consumers and vendors.

IMPLEMENTATION
• Conducted secondary research around every category involved (apparel, venue, logistics, makeup, etc), with industry sizes and trends.
• Studied consumer behaviors and how the current wedding planning platforms were being used.
• Conducted in-depth two hour-long interviews by visiting consumers ar their homes, and understanding what matters to them, and the problems they were facing.
• Successfully launched an MVP to judge if users are interested to explore. Built initial wireframes to envision user flows.

IMPACT
We pitched our learnings to the executive team and secured the funding to launch a platform starting with one category at a time.

INSIGHT
I conducted data and business analysis to arrive at the insight—users usually have varied reasons for using the platform. Hence, rather than offering flat rates/discounts, we could base discounts on user intent, i.e., a more serious user gets a lower discount and vice versa.

APPROACH
We needed to figure out the demographics and activity parameters that defined intent. For instance, an older female, living away from her home town who was also logging-in frequently would have higher intent (in the Indian context).

IMPLEMENTATION
We created a machine learning algorithm along with the data science team considering numerous such parameters, which was then integrated with the existing discounting platform (keeping all existing logics uninterrupted). Finally, we tested it with a small cohort of users, which reduced and increased discounts for users based on the algorithm. We only tested it on new users, because existing ones already have a preconceived notion towards a price point, and this could bias our results.

IMPACT
We saw a 1.3% lift in revenue (a highly significant impact, which is generally achieved by multiple projects combined).

• 3.5% — Lift in 30-day Retention Across Dormant and Churned Users

A large share of users went dormant or churned, and our re-engagement channels could not reach them: email and push depend on attention we had already lost, and the app was no longer being opened. For a B2C dating app, fewer active profiles means fewer matches for those who stay, making it a flywheel problem. Leading KPI: % of lapsed users re-engaging after outreach. Lagging: 30-day retention.

The one channel these users still opened every day was WhatsApp. Built a chatbot on WhatsApp and designed the reactivation strategy around it: instead of asking users to return to the app cold, the bot re-engaged them in conversation on a channel they already used daily, then pulled them back into the app. Unlike email or push, the chat was two-way: users could reply, and the bot could respond in the moment interest showed up. Measured 30-day retention separately for the dormant and churned cohorts, so we knew which group the lift came from.

30-day retention improved by 3.5% across dormant and churned users, the hardest cohorts to move because they had already left the product. On a $50 million ARR business, retention lifts of this size flow directly into revenue.

• 5.4% — Lift in Free-to-premium Conversion for The Test Cohort
• 0% — Change in ARPO (No Discounting Used)

Free users were not upgrading, and discounting was the obvious lever, which erodes ARPO. The real issue: free users never experienced what premium did. Our best premium feature was chatting with your matches, but a free user receiving a chat request only saw a small notification most never noticed. Leading KPI: % of free users hitting premium value in their journey. Lagging: conversion and ARPO.

The idea: surface the most compelling premium feature inside the organic user journey, contextually. When a chat request arrived, a pop-up appeared in front of the user in that moment; replying required an upgrade. Took an MVP approach: started on Android, where 70% of the user base was, and reused existing UI components from the library to keep the build small. Also started with male receivers only, a deliberate call: female users already receive overwhelming attention, and more pop-ups would have worsened their experience. Released as an A/B test for both new and existing users.

Free-to-premium conversion increased by 5.4% for the test cohort, with zero discounting and no drop in average revenue per order (ARPO). The next step was extrapolating the same contextual paywall to other cohorts and platforms.

• 12% — Improvement in Order Conversion Across Web, iOS, and Android

Users who had already decided to pay were dropping off inside the payment flow, on web and on the mobile apps. This is the most expensive leak in a funnel: acquisition, matchmaking, and plan selection had all done their job, and the loss came at the last step. Leading KPI: payment completion rate (plan selected to payment success), by platform. Lagging: order conversion and revenue.

Instrumented the payment flow step by step and platform by platform (web, iOS, Android) to see exactly where users dropped: which screen, which step, which device. Worked with the design team through low-fidelity wireframes and usability testing before committing to the build, so friction was caught early. Then redesigned the flow end to end rather than patching single screens, giving users one consistent path from plan selection to payment success on every platform. Measured order conversion against the earlier flow as the baseline.

Order conversion improved by 12% across web, iOS, and Android. This was on a product segment worth $50 million, where checkout gains translate directly into revenue rather than needing any change in traffic or pricing.

• 100% — Onboarding Funnel Efficiency, Sign-up to Activation, in 6 Months
• 98% — Increase in New ARR Over the Same 6 Months

Only 9% of business owners who signed up made it through onboarding to an activated bank account. Opening a business account requires identity checks, document review, and fraud prevention, so some friction is regulatory and cannot be removed, making every unnecessary step doubly expensive. Leading KPI: step-by-step completion from sign-up to activation. Lagging: activated accounts and ARR.

Treated it as a program, not a single test. Instrumented every step from sign-up to activation, ranked drop-offs by opportunity size, and worked through them as a continuous experimentation backlog. Established an experimentation operating model with product, engineering, growth, and compliance, so tests shipped fast without breaking regulatory or risk safeguards. The biggest single win moved document collection into the onboarding flow itself instead of over email. Built daily, weekly, and monthly dashboards to track progress and a 12-month OKR to keep every team pointed at the same outcome.

Onboarding funnel efficiency doubled in six months, from 9% to 18%, sign-up to activation. The product-led growth strategy this program anchored resulted in a 98% increase in ARR over the same six months.

• 18% — Increase in Customers Responding with Documents
• 12% — Increase in Sign-up to Account-opened Conversion

Customers were not sending the documents our fraud team needed, which blocked their bank accounts from opening. In banking onboarding, document verification is the one step you cannot remove, so the only real lever is making submission easy. Was this low user intent, a UX issue, or both? Leading KPI: % of customers responding with documents. Lagging: sign-up to account-opened conversion.

Data analytics pointed to UX, not intent. We were asking customers to switch interfaces and respond over email to upload documents. That switch added friction to an already tedious scan-and-upload task and invited distraction and procrastination. The fix carried a hard constraint: the fraud team ran its casework in Intercom, and that tooling could not change. So we brainstormed ways to collect documents inside the onboarding flow itself and launched an MVP using the Intercom widget: the customer never left the flow, the fraud team kept its tool, and the UX hypothesis got a cheap, fast test.

Customers responding with documents increased by 18%, which drove a 12% increase in sign-up to account-opened conversion. The MVP validated the hypothesis without new infrastructure or any change to the fraud team’s workflow.

• 35% — Lift in Organic Traffic

Acquisition leaned on paid and organic channels we did not control, and for a neo bank paid CAC is expensive and rented: the moment spend stops, growth stops. We wanted a defensible acquisition channel owned by the product itself, powered by customers who already trusted us. Leading KPI: % of customers sending referral invites. Lagging: organic traffic and new sign-ups.

Ran the referral program end to end: business case first, sizing what a referral channel could contribute before writing a single requirement, then discovery, execution, and post-launch iterations. In banking, referral incentives attract abuse, so every design trade-off was monitored with growth and compliance together rather than cleaned up after launch. Positioned referral inside the broader product-led growth strategy: onboarding and activation improvements meant referred users landed in a funnel that converted, and the program gave that funnel a self-sustaining source of traffic.

Organic traffic lifted by 35%. Referral gave acquisition a defensible, owned base alongside paid channels: each new activated customer became a potential source of the next one, at near-zero marginal cost.

• 12.3% — Increase in Sign-up to Paid Conversion
• 19% — increase in invoice creation (the Aha Action)
• 8% — Increase in Annual Plan Purchases
• $250,000 — ARR Added

Free users could send only one invoice, so most new users never experienced the product’s full value and never felt the need to pay. Our biggest funnel drop sat between sign-up and the first invoice, our Aha moment, and “free forever” meant no urgency to decide. Leading KPI: % of new users sending their first invoice within a week (speed to Aha). Lagging: sign-up to paid conversion and MRR.

Paired funnel data with user interviews and session recordings. Insight: freemium was working against us. A one-invoice limit meant users never saw full value, and free forever meant no reason to decide. Our question: how might we let every new user experience the full product, with a real reason to decide? Flipped acquisition from freemium to a 14-day full-access trial, a reversible A/B test on new users only. Built the trial countdown, expiry, and upgrade screens, and emails; aligned Marketing and Support on messaging; tracked sign-up to first invoice to paid; ran 6 weeks, then went to 100%.

Sign-up for paid conversion increased by 12.3%. Invoice creation, our Aha action, increased by 19%. Annual plan purchases increased by 8%. Added $250K in ARR from a six-week reversible test on new users only.

• 40% — Of Users Adopted the Voice Feature
• 70% — Of Adopters Used It on a Recurring Basis
• 60% — Reduction in Inference Cost via The Split Model Pipeline

Electricians, plumbers, landscapers: these users live in a van. The job ends and they are driving to the next site, so the invoice waits for the evening. That delay is expensive: time-to-invoice drives the cash-collection cycle, an estimate loses to whoever sends first, and pros forget job details in 2-3 days. North star: created-to-paid conversion. Leading: share of invoices started by voice.

Three design calls made the product. Autonomy: drafts never auto-send. A wrong quote hits reputation and top line, and pros promise in-person discounts that must show up; auto-send only saves one tap. Latency: my threshold was under a second, so I split the pipeline instead of using one expensive model: streaming speech-to-text so the user watches words appear, then a cheaper model with RAG turning the transcript into structured line items without hallucinating. Precision over recall: a blank price beats a guess, backed by citations (“pulled from last 3 invoices with this client”).

40% of users used the voice feature, and 70% of them used it on a recurring basis. The split pipeline cut inference cost roughly 60%. Post-launch, unsent drafts piled up, so we added an evening nudge: “you have 3 invoices waiting.”

• 72% — Reduction in Morning Load Latency
• 67% — Drop in Board-freeze Support Tickets
• 42% — Increase in Feature Adoption After AI Auto-routing Shipped

On our field service platform, the dispatch board, calendar, and routing logic shared one monolithic database with CRM and billing. Every morning, dispatchers dragging jobs locked records while technician GPS pings hammered the same system; the board slowed until operators coordinated by phone. The architecture also blocked AI auto-routing. Leading KPI: morning latency and board-freeze tickets.

I pitched leadership that this was a boundary problem, not performance tuning: scheduling is high-frequency and real-time, CRM and billing are slow-changing and need transactional precision. We decoupled scheduling and routing into a standalone microservice with its own data store. I owned the service boundary (owns assignments, time windows, routes, live GPS; references customers, terms, invoices), the API contract and event model (which system wins on conflict, what a dispatcher sees in a billing outage), and a strangler migration: reads first, writes behind a flag, account by account.

Morning load latency dropped 72%, and board-freeze support tickets fell 67%, with zero disruption to operators dispatching daily. Freed of the monolith, AI auto-routing shipped in about a quarter and drove a 42% increase in feature adoption.

• 12.8% — Lift in Bookkeeper Efficiency (Books Served per Bookkeeper)

Accounting at Bench was human-led: to grow clients, you had to grow bookkeepers. AI came in as the technology to break that constraint, and the most time-taking job was transaction categorization, so we automated that first. North star: books served per bookkeeper. Leading: suggestion acceptance and override rates. Control metric, veto over growth: wrong categories in a client’s final books.

Autonomy was tiered by confidence score: high confidence auto-categorized, lower went to senior bookkeeper review, lowest showed an editable suggestion. Precision over recall: a wrong category in final books is the error a bookkeeper defends, so the error threshold was very low. Latency was forgiving, so we used reasoning models with detailed industry-specific prompts. Evals scored accuracy separately across 50+ industries against known-correct transactions, plus a fixed pile of hard cases. Memory was the example library: senior overrides fed it every two weeks, tracking senior judgment.

Bookkeeper efficiency rose 12.8% and senior escalations on routine categorization dropped sharply. Version one hit 91% accuracy and still wasn’t trusted; a one-line “why this category” shown inline jumped trust scores 40 points. Accuracy isn’t trust.

• 8.4% — Lift in Client Retention
• 35% — Share of The User Base on Mixed-Use Accounts Served by The Feature

35% of Bench’s clients ran business and personal spending through one mixed-use bank account. Only the client knows whether a charge was business or personal, so bookkeepers chased them with questions, books waited on answers, and the friction on both sides showed up in retention. Leading KPI: % of mixed-use charges categorized by clients without a bookkeeper chase. Lagging: client retention.

The insight: this was a workflow problem between two people, not an automation problem. The bookkeeper should not guess, and the client should not answer one-off emails. So we built self-serve categorization: the ambiguous mixed-use charges went to the client, who categorized them in the product on their own time, while everything else stayed with the bookkeeper. It came directly out of the LLM auto-categorization work, where part-business, part-personal charges were the edge case no model could decide. The fix was routing the decision to the only person who actually knew the answer.

Retention lifted 8.4%, on an initiative serving 35% of the user base. The same collaborative workflow pattern, AI, and bookkeeper handling what they each know, client handling what only they know, became part of Bench’s playbook for later features.

• 8.9% — Lift in Daily Active Users
• 3.9% — Increase in Trade Volume

Investors opened the trading app to check their portfolio and left. Generic market news gave them no reason to come back: it was not about the stocks they held. For a brokerage, engagement is the business, since users who visit daily trade more. And anything shown next to someone’s money ships inside compliance constraints. Leading KPI: news consumption per user. Lagging: DAU and trade volume.

Led the hyper-personalized news engine from requirements through POC to launch on a $500M+ ARR trading product. Sourced content through partnerships with Benzinga and Seeking Alpha, owning vendor evaluation, integration requirements, and the schema normalization that made external feeds feel first-party. Personalization matched news to each investor’s own holdings. Reliability was engineered: retrieval grounded in canonical reference data so nothing was invented, offline and online evals, and latency routing so the right model served each request. First release was scoped deliberately small.

Daily active users lifted 8.9% and trade volume increased 3.9%. Both came from the same mechanism: news about the stocks you own is a reason to open the app daily, and an informed investor trades with more confidence.

• 3 Weeks to 45 Min — Customer Research Synthesis Cycle Time
• 50 — Past Reports in The Golden Eval Set Gating Every Prompt Change
• 40% — Model Cost Saved by Splitting Judge and Critique Models

The head of customer research spent three weeks per cycle reading support tickets, app reviews, and interview transcripts, then hand-assembling a deck that arrived stale and shaped by one person’s biases. A UI bug outweighed an enterprise account threatening churn because it came up more often. Hard constraint: transcripts never leave company servers. North star: synthesis cycle time, rigor held.

Shipped a multi-agent critique swarm, not a summarizer. A Synthesizer agent drafts insights; persona agents (PM, Engineering, Sales) critique the draft; a Judge forces convergence after two capped rounds, because uncapped agents debate forever. Every insight cites back to transcript timestamps and aligns to the quarter’s OKRs. PII is scrubbed locally before anything enters the on-prem vector store. Reliability is engineered: LLM-as-judge evals against a golden set of 50 past reports before any prompt change ships, per-agent token observability with a $2 auto-terminate, and checkpointed resume.

Synthesis time dropped from three weeks to 45 minutes, and rigor went up, not down: the swarm stress-tests its own logic before a human sees it. The system drafts and critiques but cannot email the CEO or touch Jira; the human keeps the final call.

Executive coaching is a luxury for the top 0.05% of professionals; the other 99.95% hope for a good boss. The timing is urgent: as AI automates functional skills, the career premium shifts to human skills, influence, communication, empathy. Senior ICs told they lack executive presence could never afford quality coaching. Vision: a world-class shadow coach in every professional’s pocket.

Built ProLeap as an ambient shadow coach, not a practice simulator. It listens to real meetings across Zoom, Meet, and Teams, analyzes them through coaching lenses (communication, leadership, influence, strategic thinking), and grounds every piece of feedback in evidence: the exact quote, the timestamp, and what to phrase differently. Privacy-first, with transcripts processed locally. The moat is the Evolution Loop: users talk back (“I prefer blunt feedback”), and corrections build a Personal Operating Manual, so the coach learns when to push and when to support. Shipped solo with Claude Code.

Shipped end to end by one person: 120,000+ lines of production code in 90 days. Currently being used by 18 users and being improved continuously through feedback.