Redapto Reads Every Support Chat You Send - So Your Team Stops Repeating Itself
Most quality teams sample under 5% of customer conversations. Redapto (YC F25) watches all of them, flags the AI hallucinations, and rewrites the training on its own.
Ask anyone who has run a customer support team about quality assurance, and you will hear a version of the same confession: they read a sliver of the conversations and hope the rest look like it. A reviewer pulls a handful of tickets a week, scores them against a rubric, and moves on. The other 95% - the frustrated refund request, the agent who guessed at a policy, the AI assistant that invented a shipping date - goes unread. Redapto, a San Francisco company in Y Combinator's Fall 2025 batch, is built around that unread pile.
The company's one-line description is deliberately plain: self-improving systems for customer support teams. In practice that means Redapto sits across a company's chat, voice, and email channels and monitors 100% of what customers and agents say to each other. It looks for quality problems, procedural gaps, and - increasingly relevant as support desks bolt on AI - hallucinations, the confident-sounding answers that happen to be wrong.
The gap it lives inCoverage first, then everything else
The 5%-versus-100% gap is the whole pitch, and it is worth sitting with. When a QA lead reviews a sample, they are not just missing conversations. They are missing the pattern that only shows up when you look at a thousand of them at once - the specific refund policy three different agents keep getting wrong, the phrasing that reliably makes an angry customer angrier. Redapto clusters complaints to find those root causes rather than treating each ticket as a one-off, and tracks sentiment across the whole book to flag where tone is quietly slipping.
Coverage alone would make Redapto a better analytics dashboard. The part the founders keep returning to is what happens after the reading. Findings do not sit in a report waiting for a manager to act on them. Redapto is designed to close the loop: it drafts updated training material, revises the knowledge base as products and policies change, and can run simulations of a support flow before it goes live - the way an engineer runs a test suite before shipping code.
How it worksA loop, not a scoreboard
It helps to picture the product as a circle rather than a chart. Conversations come in; the system evaluates them against criteria tuned to a specific brand's voice, policies, and compliance needs; it surfaces what is broken; and then it writes the fix back into the training and knowledge the team relies on. Then it watches the next batch to see whether the fix worked.
The custom-evaluation piece matters more than it sounds. A generic scorecard rates every company's support the same way. But a fintech has compliance language it legally must use, and a marketplace has a tone it has spent years building. Redapto lets teams define what good looks like for them, then holds every conversation to that bar. For a regulated business, a hallucinated answer is not a demo bug - it is a refund, a churned account, or a compliance flag, which is why catching it in real time is the feature people lean on.
The foundersEngineers who have done this at scale
Redapto was founded in 2025 by Anirudh Pupneja and Cheril Shah. Pupneja, the CEO, previously built the generative-AI platform at Coinbase, where his agent infrastructure is credited with saving the company more than $2.5 million a year. Shah's background is in the unglamorous engineering that makes AI cheap enough to run everywhere - model fine-tuning, compression, and distillation at Adobe and Amazon, work reported to have cut system latency by 20x and saved over $1 million.
That pairing - someone who has shipped production AI agents and someone who has made them efficient - reads as a deliberate answer to the problem they picked. Reading 100% of conversations in real time is only useful if it is also affordable, and support desks run on thin margins.
The marketWhere it fits
Redapto is aiming at e-commerce, fintech, and marketplace or platform companies - delivery, travel, on-demand services - where support volume is high and a single bad interaction can end a customer relationship worth far more than the ticket that ended it. The company places itself in the account-management and customer-experience software category, a market it sizes at roughly $14-16 billion in the United States.
The competitive neighborhood splits in two. On one side sit support-QA and conversation-intelligence tools that score conversations. On the other sit AI agents that answer them. Redapto's position is the seam between: it evaluates every interaction against a brand-specific bar and then acts on what it finds by rewriting the training and knowledge underneath. Whether that seam becomes its own category or gets absorbed by the players on either side is the open question - and a fair one for a company this early.
Business & backingSmall team, measurable pitch
The model is straightforward B2B software, sold on outcomes rather than seats: higher CSAT, higher net revenue retention, less churn, faster agent onboarding. In December 2025 Redapto announced a $500,000 pre-seed round led by Y Combinator, with the stated plan to deepen its work in activation, expansion, and full-lifecycle personalization while improving its data pipelines and AI accuracy. The team is still small - two people at the time of its YC listing - which is roughly what you would expect from a company whose entire thesis is that a system, not a bigger headcount, should do the improving.
For now, Redapto is a bet on a quiet idea: that the most valuable thing in a support org is not the next hire or the flashiest chatbot, but the transcript nobody has time to read - and a loop that reads it, learns from it, and closes.