Unicorn AI Summit · 21 September 2026

Build the AI OS
your company runs on

Kaushik Bhat
AI Platform, Razorpay

01RazorpayRazorpay
The ecosystem
Slash Slash
It buildsthe engineering agent
It reviewsspecialists on every pull request
It remembersa context layer agents query
It runsthe same agents in your terminal
It routesone path to every model
It servesopen models on our own hardware
02
The map

Build. Adopt. Optimise. Spread. Own.

February
Build
the core cloud agents — Discover, Coding, Reviewer
March–April
Adopt
scale across every engineering team
May–June
Optimise
evals, the gateway, open models
July–August
Spread
agents built beyond engineering
September
Own
self-hosted inference and the knowledge graph
03
Where the agent runs

Put the agent where the work already is.

WHERE WORK ALREADY HAPPENS SlackGitHubCIJira and more getting added everyday Control planeauthenticate · persist · routenever runs an agent EVERY TOOL, ONE GRANT MCP · skills · plugins Queue DATA PLANE · ONE PER TENANT Worker Agent as a subprocess file allow-list · egress default-drop unprivileged user
04
The reviewer

Review became the bottleneck. So the queue forced a specialist.

PR SecurityBugsCode qualityDesign systemi18nPre-mortem Repo-specific rules+ false-positive filter Decisionlow severity: approve 1 in 3merged withno human comment
05
The reviewer
It reviewed it. Then it approved it.
A merged pull request in razorpay/blade: rzp-slash Bot approved the changes, left an Agentic PR Review comment with status approved and four passing UI checks, and added an eligible-for-auto-approval label
06
The gateway

Nobody plans this layer. The bill does.

The agentThe reviewerAny team's agent Gateway SDKharness adaptersdeclared capabilities LLM gatewayrouting and accesscost and accounting Managed providersOpen-weight modelsSelf-hosted serving frontier, on tapmost of the volumeour own hardware
07
Evals

Changing a model should be a measurement, not a migration.

STATICCONTINUOUSDYNAMIC RUBRICS A frozen slate,every nightSampled fromlive trafficCriteria writtenper task family did this change make it worse?is it broken right now?what does good mean here?
08
Slash economics

Usage keeps rising. The bill keeps falling.

spend peaks · routing, evals, open models AI work done cost per task SIX MONTHS AGOTODAY
09
Context

Curate knowledge. Index code. Query live state.

Curated knowledgeCode truthLive state reviewed like codeindexed nightlyqueried, never indexed Knowledge agentone MCP endpoint Slash and other agentsEngineer's LaptopsAny MCP client
10
Local Harness

Everything Slash learned, on every laptop.

WHAT THE PLATFORM ALREADY HAD Certified skillsMCP catalogueSlash DiscoverReviewer's rules Slash Harnessconfigures the agent you run The coding agent on your laptopunchanged, now carrying all of it AND THE STANDARD LOOP IT INSTALLS discoverbrainstormplanplan reviewtest firstsimplifystaged review
11
Our AI OS

Own the floor. Then trust the models.

WORK ARRIVES Slack Github CI · Tickets · Events APPLICATIONS Slash Slash Reviewer Team Specific Agents RUNTIME identity · tools · queue CONTEXT & EVALS Slash Discover Slash Evals MODEL ACCESS Slash Gateway · Gateway SDK · Model Auto-router MODELS & SERVING frontier · open-weight our inference engine
12
Platform

Other teams started building.

Slash50 other agents built org-wide
13

So: what will your company run on?

14RazorpayRazorpay