Year
[Add year]
Disciplines
Accuracy
extraction accuracy across 100K+ documents a month
Efficiency
lower compute cost than an LLM-only approach
Scale
agreements labelled by domain experts over two years
What were people actually doing?
Reading by hand. A single question about one deal meant opening several documents, scanning hundreds of pages, and copying what mattered into a spreadsheet. Excel was the source of truth because nothing else could be trusted to be one.
The old way
What the old way lacked
Connection
Findings lived in one person’s spreadsheet. Nothing moved between systems without a human carrying it.
Context
By the time a comparison was assembled, the underlying terms had often moved on.
Traceability
An answer without a source is unusable to a lawyer. Most answers arrived without one.
Speed
Work that decided millions took days, and the delay itself shaped which questions got asked at all.
How does this make people feel?
Outnumbered
Users described being buried. The volume was not just slow, it changed their behaviour: they stopped asking questions they could not afford the time to answer.
And the stakes are not abstract
These are contractors, lawyers and project financiers. A misread term carries legal and financial exposure, so any tool that answers quickly but cannot show its work is worse than no tool at all.
What can we do?
Form a hypothesis
If we give people a single place to ask questions of every document, and anchor every answer to its source, then we remove the manual search burden without asking them to trust a black box.
Align on the questions that drive us
How might we create a world where finding an answer takes seconds instead of days?
How might we create a world where every answer carries its source with it?
How might we create a world where people can compare deals without losing their place?
How might we create a world where the system admits what it does not know?
Define success
Every answer traceable to a source document
Cross-deal comparison without leaving the page
Answers returned in seconds, at 95% extraction accuracy
Experiment
A new mental model
Rather than rebuilding the dashboard, I fixed the one seam where friction compounded: context. Filters persist across views, breadcrumbs anchor position, and deals compare inline, so the path forward never doubles back.
Filters hold
Compare inline
Decide
Context never resets
Progressive disclosure for the chatbot
The chatbot became the platform’s anchor, so it had to feel systematic on the first try. I split it into three stages so users start broad and narrow in gradually instead of facing a blank box.
01
Welcome view
Entry points organised into topic-based buckets, so users land somewhere that tells them where to start.
02
Intermediate view
Context-aware follow-up prompts carry the conversation forward.
03
Answer view
Anchored in a data grid rather than prose, because these users already treat Excel as their source of truth.
Before and after
Before
After
Before state
Nested views. Every comparison was a round trip, and every return reset the filters.
Launch
The redesigned dashboard and the chatbot shipped as one system. The same persistent filter bar, breadcrumb trail and comparison pattern carry across Agreements, Datafinder, Models and Find & Compare Terms, so learning one surface teaches you all of them.
Agreements
Deal cards with source sentence and named parties
Datafinder
Key terms traced back to their contract
Models
Financial models with IRR, Capex and O&M rollups
Built to survive an enterprise
Data sovereignty — in-region hosting on AWS Saudi Arabia to meet local data regulations.
Role-based access control — different groups, different visibility into sensitive deal data.
Verifiable answers — every response traceable to a source document.
Concrete use cases — grounded in water and electricity distribution projects the client would recognise.
What difference did we make?
The same numbers as the top of this page. They mean more now.
Accuracy
95%
extraction accuracy across 100K+ documents a month
Efficiency
90%
lower compute cost than an LLM-only approach
Scale
12,000
agreements labelled by domain experts over two years
Going further
This launch had real impact, but all software is an unfinished experiment and we got things wrong. Once real usage started I tracked and specified the interaction bugs that were quietly eroding trust: feedback controls that scrolled users back to the top of a response, sticky navigation that handed off too early, and a data table that intermittently showed no results until you toggled views, or showed results belonging to a different question. Specifying these precisely, with reproduction steps rather than ‘the table is broken’, was part of protecting trust in a product whose entire premise is giving people reliable answers.
What’s next
A feedback modal on negative ratings, turning criticism into structured data for improving the model.
Autocomplete in the welcome view, helping users complete prompts as they type.
More Projects
Build with Intention
Get in touch!
© 2026 Nidhi B Gowda
Mobile: (914)-308-6701