# Document search engineers can verify

> Document search for a New York infrastructure company, where every answer points to the exact page and paragraph it came from.

- Client: TCE, a New York infrastructure company
- Type: Enterprise AI document search, retrieval with source citations
- Category: Enterprise AI
- Duration: About 3 months to build
- Status: In production and still used
- Stack: Azure-aligned deployment · PDF ingestion · Retrieval with citations · Role, team and project permissions
- Role: End to end, first call to launch
- Result: ~83% estimated reduction in document-search time (Based on engineer interviews and client estimates, not formal usage logs or a controlled pilot.)
- Result: 60 → 10 estimated minutes searching per engineer per working day (Includes time spent reading AI answers. Engineer interviews and client estimates.)
- Result: Live in production and still used by TCE (Status in October 2026. No active-user count is published.)
- Adamant Code case study: https://adamantcode.com/case-studies/tce-document-intelligence
- Built by: Luka Abramović (https://lukaabramovic.com/)
- Canonical: https://lukaabramovic.com/work/tce-document-search/

---

## The starting point

TCE engineers review dense bid-document packages for large infrastructure projects. A single package can span 5 to 20 PDFs and roughly 5,000 to 50,000 pages. Finding a requirement meant searching PDFs one at a time with Command-F, working across local files, or going back to printed pages and highlighted notes.

## A search problem with a trust requirement.

A plausible AI answer was not enough. Engineers had to verify every response against the original documents, and the system had to respect how TCE's projects, teams and permissions already worked.

### Scale

Thousands of technical pages per bid package.

### Evidence

Evidence over black-box output: citations, page references, exact paragraphs and highlighted source text on every answer.

### Access

Enterprise access rules for admins, managers, users, teams and projects.

## My role

I ran this end to end, from the first conversation through discovery with TCE's bid-review engineers, scope, solution design and leading the build with the Adamant Code team, to launch. The central product decision was mine to make early: help engineers reach the evidence faster, and leave the engineering judgment that follows with them.

1. First call
2. Discovery
3. Scope & proposal
4. Solution design
5. Build lead
6. Launch
7. Iteration

<StepDiagram
  label="Workflow diagram: an engineer's question goes to retrieval across the project's documents, which builds a cited answer, which the engineer reviews against the source"
  steps={[
    { label: 'Question', detail: 'Plain language, no keyword hunting' },
    { label: 'Retrieval', detail: "Relevant passages from the project's document set, within the user's permissions" },
    { label: 'Cited answer', detail: 'Page, paragraph and highlighted passage', tone: 'accent' },
    { label: 'Engineer reviews', detail: 'Judgment stays with the person responsible', tone: 'dashed' },
  ]}
>
The AI brings engineers to the evidence. The decision that follows stays with them.
</StepDiagram>

## From question to verifiable answer.

1. Ask in natural language, without translating the question into keywords or opening every PDF.
2. The system retrieves the relevant passages from the project's document set and builds an answer from them.
3. Citations link the answer to its page, paragraph and highlighted passage.
4. The engineer reviews the evidence in context and decides how it informs the work.

## More than an answer box.

### Document and project management

Packages organised by project, so each search runs against the right set of documents.

### Source-cited search and summaries

Answers and summaries with citations, page references and the exact source paragraph highlighted.

### Team and role permissions

Admin, document, project, team and role permissions that follow how TCE already works.

### A deployment around TCE's environment

An approach designed around TCE's Azure environment and sensitive enterprise data.

## In production, still used.

The system is in production and still used by TCE. Based on engineer interviews and client estimates rather than formal usage logs or a controlled pilot, average document-search time fell from around 60 minutes to around 10 minutes per engineer per working day, an estimated 83% reduction, with the original evidence still in view.

<MetricTiles
  size="m"
  items={[
    { value: '~83%', note: 'Estimated reduction. Engineer interviews and client estimates, not usage logs or a controlled pilot.' },
    { value: '60 → 10', note: 'Estimated minutes per engineer per working day, including time spent reading AI answers.' },
  ]}
/>

> The enterprise chatbot can answer that question and provide multiple sources in an instant with highlighted sources. So that's been a huge value for our company!
>
> – Patrick Besser, Data & Systems Engineer, TC Electric

## What this proves

For expert teams, useful AI is not the system that sounds most certain. It is the one that gets people to the right evidence faster, shows where every answer came from and keeps judgment in human hands.