Research intelligence from banner tables to decks

The problem

You already have the numbers in multi sheet banner files, but turning them into a 30 to 60 slide client deck still means manual copy, easy rounding errors, and uneven narrative quality. Junior staff spend days on layout while reviewers cannot see which cells fed which claim.

You also need a way to steer the analysis. Some runs should stay descriptive, others should add forward looking or action oriented angles. That control has to sit next to the grid, not in a separate ticket to a design team.

What we built

We shipped a Next.js workspace paired with a FastAPI service. You upload an Excel banner in standard extensions. The backend parses headers, sheets, and data rows, then returns a structured grid the UI renders with per cell labels and statistical flags so you can spot higher and lower readings before you commit them to the deck.

You select cells directly on the grid, or you run auto select which targets dozens of cells sized for a medium deck and prefers statistically marked cells first, then spreads the rest across rows and column groups so the deck is not dominated by one block of the table.

• An analysis panel lists chosen data points, supports clear and reset, and records optional analyst notes plus insight mode toggles for descriptive, predictive, and prescriptive emphasis.
• When you generate, the client stores the selection payload in session storage and opens a generate view that posts to a streaming endpoint and reads server sent events in real time.
• The server pipeline walks staged work in a fixed order. It builds a semantic model from the banner, mines insights with a large language model, drafts a story plan for the deck, then validates that plan so numeric claims trace back to real cells instead of model drift.
• Chart slides render through a chart generator that writes image assets. Image slides can use a hosted image model for backgrounds and diagrams where you enabled that path.
• Slide assembly runs generated presentation code through a Node based PptxGenJS runner so the output is a real PPTX file. A conversion step produces a PDF preview you can open in the browser while the timeline shows each step starting, streaming chunks where applicable, and completing or failing with a visible message.

What you gain

You shorten the path from locked data to reviewable slides. Stakeholders see the same story plan and preview URLs update as the run advances, which makes remote sign off easier than passing binary attachments in chat.

The validation stage is the guardrail. It forces numbers on slides to reconcile with the workbook before you present to a client, which is where most manual processes fail today.