← Back

Govr System
Build

An AI research co-pilot that prefers signal over volume.

Deployment
Live MVP
Role
Solo
Low-Grade Citations
62%
Eval Set Size
180
Time to MVP
7 wks
Product Spec RAG Architecture Eval Design Zero-to-One

Govr is a zero-to-one AI system I designed and shipped solo — a retrieval-augmented workflow that prioritizes source quality and evidentiary weight over page count. Built to help analysts defend a thesis, not drown in tokens.

Generic LLM research tools optimize for coverage. Analysts don't need more text; they need fewer, better citations. The problem is a product one: how do you design a loop that rejects low-quality retrieval even when it's cheap?

01
Wrote the product spec before touching a model — defined what 'quality' meant across five source types.
02
Architected a multi-stage retrieval: breadth pass → source-grade re-ranker → evidence synthesis.
03
Built the eval harness first. No feature shipped without a regression on a hand-labeled set.
04
Kept the UI deliberately text-first — the product is the reasoning trail, not the chat.
MVP live and used by a small group of analysts and founders.
Source-quality re-ranker cut low-grade citations by ~62% against baseline.
Shipped end-to-end as sole PM, architect and builder.
Next Case
Distribution & GTM
The IHCL Strategy →
← All Work
© 2026 Dhanishta — All thinking, some shipping.
Open to Roles
☽☀