Sathvik VK

Perfstaq

Business software built so an AI assistant can use it safely: every tool runs through one pipeline, and risky calls wait for a person.

Status
In development. The platform is under active development; parts of Perfstaq are live on perfstaq.com subdomains
My role
Founder. Product scope, system design and most of the platform code.
When
Aug 2026 to present
Context
Perfstaq Private Limited
Capabilities
LLM apps, Agents, Memory, Web, Integrations
An approval screen: an assistant asks to send a mutual NDA, with deal details and Approve and Reject buttons.
An assistant's request waiting for approval. Local build with seeded demo data; names and companies are fictional.Screenshot

What it is

Perfstaq has had two product lines so far, and it is worth keeping them apart.

The first was an AI marketing product: a Brand Brain that holds what a brand is, campaign planning on top of it, generated creative, and a human review step before anything is used. That was team work. I led discovery and product direction and contributed code; I was not the main committer.

The second is the platform, which is mostly my code. One signup gives a company its own workspace with apps for e-signatures, boards, sales, calendar, notes and finance, and one MCP connector so Claude, ChatGPT or Cursor can operate them.

What I built

On the platform I designed the architecture and wrote most of it.

  • A single tool pipeline. Every capability is defined once and runs through the same steps: validate input, check the app is enabled, check role and plan limits, hold risky assistant calls for approval, run, then write one audit record. The web UI and the MCP gateway both go through it.
  • The platform's own OAuth server for MCP clients, with per-token rate limits stored in Postgres.
  • Host-based multi-tenancy, so each organisation works on its own subdomain.
  • The app modules: Sign, Board, Sales, Calendar, Notes and Finance.
  • Plain SQL migrations applied under an advisory lock, and a test setup with unit, browser and accessibility checks.

Decisions

The pipeline exists because of one worry: an assistant with direct access to business tools will eventually do something expensive. Routing the UI and the assistant through the same checks means there is no second, weaker path to forget about.

Approvals are data, not prompts. A risky call from an assistant becomes a pending record that a person approves or rejects in the product, and it expires if nobody does.

How it fits together

Callers
Web app
server actions
AI assistant
MCP over OAuth
Pipeline
Validate
typed input
Authorise
app enabled, role, plan limit
Approval
risky assistant calls wait
Apps
Sign
Board
Sales
Calendar
Notes
Finance
Record
Audit event
one row per call, in Postgres
A person in the web app and an assistant over MCP reach the same tools through the same checks. Simplified diagram drawn for this page, not a product screenshot.

Built with

Next.js 16React 19TypeScriptPostgresDrizzleMCP SDKOAuth 2ZodTailwindVitestPlaywrightCloud Run

Screens

Next case studyOr press the right arrow keyThaaliA calorie tracker for Indian meals. Photograph a plate and get an editable list of dishes with calories. Launched as a PWA at thaali.life.