[ INVESTMENT_RESEARCH ]

AI Research Report & Deck Generator

2025Data Pipelines & AutomationIn-depth

What it is

Source documents in; a structured research report or presentation out.

Client work is shown without identifying imagery

[ THE_PROBLEM ]

Why this existed

Turning a data room into a first-draft research report takes an analyst a week, and most of that week is reading and restructuring rather than thinking.

[ WHAT_WE_BUILT ]

What we built

A Python service handles ingestion — per-report vector stores built from uploaded PDFs, with a search layer over them. The application side plans the report, generates section content grounded in the retrieved documents, converts documents between formats, produces presentation slides from generated content, and stores every report with a history and detail view. Auth, a saved-report library and progress indicators over what is a long-running generation job.

Python ingestion service building per-report vector stores from uploaded PDFs
Search layer over the indexed documents
Report planning followed by grounded section generation
Document format conversion
Presentation slides produced from generated content
Report storage with history and detail views
Authentication, saved library and progress indicators for long-running jobs

[ HOW_IT_IS_USED ]

How a company uses it

An investment or consulting team turns a data room into a first-draft report in an afternoon. The analyst edits rather than writes — which is a different job and a far faster one.

Built with

Next.jsFastAPIFAISSAnthropicPineconeMongoDB

[ COMMON_QUESTIONS ]

Questions clients ask

Why separate planning from generation?

Because generating a long document in one pass produces something that drifts and repeats. Planning the structure first, then generating each section against retrieved sources, keeps the report coherent and makes each section independently checkable.

Why does the progress indicator matter?

Because generation takes minutes, not seconds. Without visible progress users assume it has hung and refresh, which is how you get duplicate jobs and abandoned sessions.

Does the analyst still need to review it?

Entirely. The gain is that they edit rather than write — a different and much faster task. Anyone shipping generated research unreviewed is taking a risk that will eventually be expensive.

[ RELATED_WORK ]