Inspiration

Investment teams usually know which metrics matter, but every new deal presents the evidence differently. Revenue may appear across financial models, pitch decks, forecasts, third-party databases, and management reports. Analysts spend significant time locating the right numbers, reconciling definitions, rebuilding calculations, and documenting where each conclusion came from.

We built Ingenium to turn an investment firm’s existing diligence methodology into a reusable semantic layer and metric catalog. Our goal is to help analysts reach answers faster while keeping every number, formula, assumption, and source open for review.

What it does

Ingenium is an evidence-first data agent for investment teams.

For each new project, Ingenium organizes the company’s data room, identifies relevant evidence, and applies the investment firm’s predefined metric definitions. Initial catalog metrics include:

  • Five-year total revenue, combining compatible Actual and Forecast periods without double counting
  • Average annual revenue growth based on the firm’s catalog definition
  • Valuation, using an explicit sourced value or a revenue-multiple scenario

Every result includes its formula, selected inputs, periods, Actual or Forecast status, assumptions, and source locations. When required evidence is missing, Ingenium returns an explicit data gap instead of inventing a number.

Users can inspect calculation lineage, review conflicting evidence, generate a source-linked diligence report, preserve human revisions, and export the result as PDF, DOCX, or Markdown.

How we built it

We built Ingenium as a connected evidence workflow:

  1. Source documents and company information enter a project-specific data room.
  2. The extraction pipeline converts PDF and Office documents into source-bound claims with file hashes and page or field locations.
  3. Structured observations and calculations are stored in Firebase, with DuckDB supporting import and offline compatibility.
  4. The backend matches evidence to metric definitions and records calculations using exact decimal values.
  5. The React and TypeScript frontend displays dashboards, calculation lineage, evidence states, reports, and citations.

We used Akta as a company data source, preserving its provider and field provenance in the project evidence. The demo uses prepared Akta data so the presentation does not depend on an unstable live connection.

We used the Nicrron API for model calls in the backend, including generating calculation code, drafting cited report content, and answering company questions against the stored evidence catalog.

We used PRISM to create the System Map, helping us organize and communicate how sources, storage, agent calculations, evidence lineage, and the user interface work together.

Challenges we ran into

The hardest challenge was preserving meaning while combining data from different sources. Two values labeled “revenue” may refer to different periods, entities, definitions, currencies, or scenarios.

We also needed to prevent the model from inventing calculations or citations. Generated calculations therefore run through a restricted backend workflow, and every saved result keeps its input IDs, generated code, execution record, assumptions, and source references.

Another challenge was connecting document extraction, structured storage, calculations, and report generation without creating competing data models. We solved this by maintaining clear contracts between extraction, calculation, and frontend components.

Accomplishments that we're proud of

We created an end-to-end prototype that connects source evidence to an investment report.

Our system can:

  • Preserve exact source locations and document hashes
  • Generate and store auditable calculations
  • Keep Actual, Forecast, and assumptions distinguishable
  • Show conflicting or missing evidence without choosing an unsupported answer
  • Link report statements back to original quotations
  • Preserve human edits separately from agent-generated content
  • Export a source-linked diligence report

We are especially proud that Ingenium can explicitly refuse a calculation when evidence is insufficient. This makes the product more trustworthy than a system that always produces an answer.

What we learned

We learned that trust in financial AI depends less on how confidently it writes and more on whether users can inspect its work.

A strong diligence system must preserve definitions, periods, assumptions, and provenance throughout the entire workflow. We also learned that the semantic layer should belong to the investment firm. Different firms may calculate or label the same metric differently, so the system must version and disclose the applied definition rather than silently imposing a universal formula.

Finally, we learned that AI works best in this context when it supports human judgment instead of replacing it.

What's next for Ingenium

Our next step is to turn the current calculation workspace into an automatic Catalog Runner. When users create a project, Ingenium will evaluate the complete metric catalog against the project’s data room, calculate supported metrics, and identify missing inputs automatically.

We also plan to add:

  • A Catalog administration and versioning interface
  • Automated processing for newly uploaded documents
  • Expanded external data-source connections
  • Production authentication and project permissions
  • Shared team review and approval workflows
  • More configurable valuation and return scenarios

Our long-term goal is to make an investment firm’s diligence methodology reusable across every deal while keeping the evidence attached to every conclusion.

Built With

  • akta
  • fairbase
  • gpt
  • nicrron
  • runtable
Share this project:

Updates

Submission history