RAMP

Ramp helps Damm UK account managers prepare scheduled customer calls with generated forecasts, custom reports, and clear visual support for gaps, wins, drivers, promos, and next actions suggestions.

Comment

Inspiration Damm's UK account managers walk into customer calls with a target, a forecast, and a gut feeling — and very little that connects the three. We wanted the moment before the call, when an AM is asking "what should I push, and what should I defend?", to be powered by the same data the brewer already has sitting in parquet files. RAMP is built around Sarah, a fictional AM whose Monday morning we tried to make 30 minutes shorter.

What it does RAMP forecasts UK beer demand 9 months out for every SKU × sub-channel, flags which conversations each customer needs that week, and recommends a sized commercial action — multi-buy, brand push, channel focus — that will actually close the gap. Every recommendation is grounded in a visible data source: historical promo ROI, the calendar, or the target shortfall itself. A what-if simulator lets the AM stress-test the play before committing, with honest £ ROI math instead of vanity uplift.

How we built it The backend is FastAPI + Polars driving a four-model ensemble — LightGBM quantile, AutoARIMA, Amazon's Chronos zero-shot, and a hand-rolled CMBC carve-out — blended per channel because grocery and on-trade behave nothing alike. A post-forecast layer applies per-SKU seasonality and event boosts (World Cup, Bonfire Night, etc.) calibrated against actuals, and SHAP surfaces why each forecast moved. The frontend is Next.js 16 + React 19 + Tailwind, with an LLM brief layer that falls back to deterministic prose so the page never breaks on a credit-out.

Challenges we ran into Our biggest fight was forecast flatness — the ensemble was right on average but visually dull, which made every demo screenshot look broken. We also burned a day on a simulator that gleefully showed "+£200k uplift" with no discount cost; the fix was modeling a 55% promo participation rate so the £ math stopped lying. And the gap-closer recommendation initially solved for the cumulative shortfall, which let it claim "promo done" while half the months still missed target — we rewrote it to solve for the worst month.

Accomplishments that we're proud of The Decision page reads like a one-pager a real AM could walk into a Trolley King meeting with — chart, plays, ROI, external context, all in one screen with the customer name in the chip so the scope is never ambiguous. The "Close the gap" play is genuinely useful: it inverts the simulator's lift curve to size the cheapest discount that clears every at-risk month's target, and tells you honestly when it can't. And the whole product degrades gracefully — LLM down, no problem; missing externals, no problem; the demo still runs.

What we learned Forecasting accuracy alone doesn't sell — surfacing why the model said what it said (SHAP, weather, search trends, events) is what turns a number into a decision. Honest negative cases ("4 of 7 months still short") build more trust than aspirational positive ones, especially with a commercial audience who's been burned by every dashboard before. And shipping behind a feature flag pattern — measure MAPE, ship if better, roll back if worse — saved us at least twice from "improvements" that quietly regressed the model.

What's next for RAMP Weekly granularity for promo planning, since most UK trade calendars run W-by-W not month-by-month. A "what closed last quarter" loop so realised lift feeds back into the historical-ROI table the gap-closer reads from. And a multi-AM view for regional managers — RAMP today is one persona deep, but the same data model already supports rolling it up.

Built With

  • fastapi
  • huggingface
  • next.js
  • python
Share this project:

Updates