Collections, rebuilt as infrastructure.

Every rupee a bank lends is a rupee it eventually has to collect. In India that recovery happens through a sprawling, offline network of collection agencies, field agents, and tele-callers that no lender can fully see and no regulator can easily audit. YuCollect set out to turn that network into infrastructure: one place where lenders discover and rank compliant agencies, allocate accounts, and watch performance down to the agent, and where agencies get the tools to be discovered, stay compliant, and collect digitally.

I led design across the arc. The hard part was never a screen. It was sequencing an enormous, multi-sided system into versions that each earned the next: an MVP that only had to prove both sides would show up, a Version 1 that made the marketplace trustworthy enough to run real money through, and a Version 2 that made it intelligent. This is that story, in order.

India's credit market exploded over the last decade, but the machinery for recovering that credit stayed stuck in the analog era. When a loan goes overdue, a lender hands the account to a debt collection agency, of which there are thousands, most small, regional, and unrated. The agency hires field agents and tele-callers, dispatches them, and reports back on paper or over the phone. The lender has almost no live view of who is calling their customer, what is being said, whether the agent is DRA certified, or whether a complaint is brewing that could turn into a regulatory problem.

This opacity is expensive on every axis. Lenders overpay because they cannot compare agencies on price or performance. Recovery suffers because allocations go to whoever is familiar, not whoever is best in that pincode. And compliance risk sits unmanaged, in an industry the regulator watches closely because it touches vulnerable borrowers. YuCollect, built inside Yubi, the debt marketplace founded in 2020 by Gaurav Kumar and backed by Peak XV, Lightspeed, and Dragoneer, was the attempt to give this whole layer a nervous system.

The reason collections had resisted productisation is that it is not a single workflow. It is a marketplace with five parties who each experience a different failure, and any product has to fix all of them at once or none of them adopt it. That is the core design complexity of YuCollect, and it shaped every version decision that followed.

A two-sided marketplace already has a cold-start problem: neither side joins until the other is there. YuCollect had a five-sided version of it, wrapped in regulation, on top of an industry that had never used software. That is why it could not launch as the finished eight-module platform it is today. It had to be sequenced.

Designing for a marketplace means designing for cohorts whose incentives point in different directions. A large private bank and a two-person regional agency both live on YuCollect, but almost nothing about their needs, literacy, or trust threshold is the same. These are the cohorts that drove the interface decisions.

Put every role on one map and the system's shape appears. Allocation flows downward, from the lender's search to the owner's acceptance to the manager's plan to the agents' calls and visits. And the truth flows back up the same rails: a borrower's payment rolls up through the manager's reconciliation, the owner's compliance submissions, the reviewers' approvals, into the score the lender reads before allocating again. The flywheel at the top is the whole business: a good score earns the next allocation.

A journey map shows what each role does. It does not show what the work feels like, and in collections the feelings are the product constraints: fear of audits, fear of harassment accusations, the fatigue of dialing blind. These maps, drawn from the interviews behind the compliance and CRM work, are what each role thinks, feels, does, and suffers, and the last row of each is what the design answered with.

The MVP had exactly one job: prove that lenders would look for agencies here, and agencies would want to be found here. Everything else was deliberately cut. The bet was the Discovery Hub: a pincode-level directory where a lender could search 19,000+ pincodes and see agencies ranked, rated, and reviewed on real performance rather than reputation, then allocate accounts in bulk.

The single hardest design constraint showed up immediately. Lenders will not upload a portfolio of overdue customers, some of the most sensitive data a bank holds, into a marketplace where agencies might see it. So the MVP was built around a zero-data-sharing allocation engine: agencies are matched and ranked, and accounts are allocated, without exposing borrower data across the marketplace. Solving that trust problem first is what made everything after it possible.

Design-wise, the MVP lived or died on two moments: a lender's first search feeling instantly more legible than the spreadsheets and phone calls it replaced, and an agency's onboarding being simple enough that a small regional shop could get listed without hand-holding. Discovery had to look less like a database and more like a decision. Get an agency ranked, priced, and allocated in one sitting, and both sides had a reason to come back.

The most dangerous moment in the whole marketplace was a registration form. Supply-side growth depended on thousands of small agencies, many of which had never used business software, filling out company details, owner details, coverage, languages, and compliance documents. A plain form here is where two-sided platforms quietly die. I had seen this exact failure before, at Khatabook, where a KYU form converted under 2% until we replaced the form-feeling with a visual reward, and 17,000 merchants signed up in two weeks. YuCollect's onboarding was designed on the same theory, rebuilt for a B2B audience.

The theory: every unit of effort a user gives must return a visible unit of identity. People abandon forms because effort flows one way, field after field with nothing coming back. So in YuCollect's signup, nothing is asked without something appearing. Type your company name, and your agency's office materialises beside the form with your name painted on the signboard. Declare yourself the owner, and the owner steps out in front of the building while a business card forms with your name on it. Add your pincode, and the street sign plants your agency in its city. By the time the form is done, you have not filled in a database. You have built your firm's presence, and it is standing there looking back at you.

The mechanic runs on four principles, each doing a specific job. Together they turn the highest-drop-off moment of the funnel into the moment the product first proves its promise.

The same theory then carries into the deeper onboarding, where the stakes rise from a signup to compliance documents. There the artifact under construction is the marketplace profile card itself: every section completed adds a real, lender-facing attribute to the card beside the form, cities, agents, years, languages. You are not doing paperwork. You are watching the thing lenders will judge you by get stronger with every field.

Screens from the YuCollect onboarding revamp design files. Hypotheses are stated as designed, with the Khatabook reward programme as the precedent evidence.

The MVP proved liquidity. Version 1 had to make the marketplace safe to run at scale, because a bank will not move a real portfolio through a platform it cannot govern, and an agency cannot grow on allocations alone if it has no way to operate them. So V1 added the two things a regulated marketplace needs most, trust and tooling, through three product lines.

V1 is where the product stopped being a directory and became an operating system. The design challenge shifted from persuasion to density: how to give a large agency the depth and bulk controls it needs, and a small one a path through the same screens without drowning. Progressive disclosure, role-aware defaults, and one consistent system across every module were what held it together.

One module from Version 1 deserves the full story, because it shows how the platform actually got built: small team, hard constraint, research-led. Compliance Plus was designed by a team of one product manager, one designer, and three engineers, on a three-week clock, with a launch lender already waiting. Lenders had started asking for exactly this during sales demos, so the feature was not just a usability fix. It was the thing standing between the platform and its next signed bank.

The ritual it replaced ran on Excel, email, and WhatsApp. Every lender requires its agencies to keep dozens of compliance documents current, police verifications, DRA certifications, service agreements, purging letters, and the chain of custody for those documents was three people forwarding files at each other with no shared state.

Three weeks did not allow for guesswork, so the design ran on four research insights from interviews with lenders and agency owners, and each one became a specific product decision.

It shipped in the three weeks, and it landed: the launch bank partnership closed, compliance work moved out of Excel, email, and WhatsApp into one place, and even in-person agency visits now log into the same timeline as digital submissions, one auditable record for the regulator.

And one honest result, because real case studies have one: the assumption that a visible compliance score would motivate agencies through lender visibility did not work as expected. The score made compliance legible, and lenders use it daily to rank their agencies, but it did not on its own change agency submission behaviour. Motivation needed a stronger loop than a number, which is exactly the lesson the visual-reward onboarding was built on.

Everything above is infrastructure. This is where it lands: a tele-caller with a headset, a campaign of twenty-four thousand overdue accounts, and a target for the month. The CRM is the screen that agent lives in all day, and it was designed around one conviction: the agent should never have to decide who to call next, hunt for context mid-call, or describe what happened in free text. The system triages, briefs, and records. The agent talks.

The worklist opens pre-triaged into outcome buckets, paid, promise-to-pay, follow-ups, not contactable, each carrying its record count and outstanding value, with the campaign's collected-versus-remaining total always in view. The queue is the strategy: the agent works down a list that is already ordered by what matters, one tap from row to live call.

Tap a row and the borrower workspace assembles the entire relationship on one screen. The live-call header keeps in-call actions, WhatsApp, SMS, disposition, end, one reach away. Account financials, payment history, the last captured field location, and the best and last dispositions sit beside a full activity timeline: every call with its recording, every field visit with photo proof, every payment link sent, every supervisor reassignment of channel or agent. It is the platform's one-shared-record idea, zoomed all the way in to a single borrower.

The most consequential design decision is the smallest-looking one: no call ends without a structured disposition. The right rail offers a coded taxonomy, callback, not contactable, promise-to-pay, refuse-to-pay, with sub-dispositions, labels, and comments. That closing tap is what feeds everything upstream, the lender's channel-wise dashboards, the compliance trail, and eventually the propensity models of Version 2. The agent thinks they are wrapping up a call. They are actually writing the marketplace's ground truth.

The same discipline serves every cohort from section 03 at once. The agent gets a day that runs itself. The agency owner gets channel planning and reassignment, moving an account from call to field when the phone stops working. The lender gets dashboards built from coded outcomes instead of hearsay. And the borrower gets contact that is logged, recorded, and reviewable, which in this industry is the difference between recovery and harassment. One screen, four constituencies, no free text.

By Version 2 the marketplace worked and was trusted. The frontier moved from infrastructure to intelligence: recover more, from the right people, on the cheapest channel that works, and give agencies the fuel to grow. Four product lines pushed YuCollect from a place you run collections to a place that makes collections better.

The through-line across all three versions is that YuCollect grew the way infrastructure has to: liquidity first, then trust, then intelligence. Each layer was only buildable because the one under it existed. You cannot sell propensity scoring to a lender who does not trust the marketplace, and they will not trust a marketplace with no liquidity to trust.

Underneath the versions is a single idea: put every party around one shared, transparent record of each allocation. The lender, the agency, the agent, and the compliance trail all read the same truth, so nobody has to reconcile five versions of what happened.

The compounding is the point. Discovery created liquidity. Compliance and portfolio tooling turned that liquidity into trust, which is what unlocked large regulated lenders. Intelligence and digital channels then drove the cost of a recovery down by more than half. Each version was a different design problem, persuasion, then governance, then orchestration, but they add up to one thing: the collections layer of Indian credit, finally legible.

Headline figures are YuCollect and Yubi published results. Version-level sequencing reflects the product build as I led it.

The lesson I carry from YuCollect is that in a multi-sided, regulated system, sequencing is the design. The eight-module platform was always visible on the roadmap, but shipping it in one go would have collapsed under its own cold-start. What worked was ruthless order: prove liquidity with the smallest possible marketplace, earn trust with compliance and visibility, then, and only then, layer on intelligence. Every screen mattered, but the decision that mattered most was what not to design yet.