CUT TO:
INT. PROJECT ARCHIVE — STORYBOARD ROOM
The USER opens Ambulance Demand Forecasting Platform.
DIVAKAR DESSAI
Let's run through the shots.
CASE FILE / Machine Learning / Forecasting / Full Stack
Ambulance Demand Forecasting PlatformPredicting When and Where Demand Happens
Built a full-stack ML platform for forecasting ambulance demand by location and time. Combined Poisson XGBoost forecasting with probabilistic confidence metrics and an interactive React dashboard for operational planning.
2025–2026
01–02 / OPENING SEQUENCE
Establishing the world
01 / Establishing Shot
The Problem
Ambulance demand varies significantly across both location and time. Historical call records can reveal these patterns, but raw trip data alone does not tell operators how much demand to expect tomorrow, which locations may become busy, or how confident a forecast should be.
02 / Wide Shot
The Context
The platform was developed during a software engineering internship to turn historical ambulance call data into operational forecasts. Rather than producing only a single predicted number, the system was designed to communicate both expected demand and the uncertainty surrounding each prediction.
03 / CHARACTER NOTE
DIVAKAR'S ROLE
Subject
Divakar Dessai
Production
Ambulance Demand Forecasting Platform
Take
03 / Role
Role notes
DIVAKAR DESSAI
I developed the end-to-end forecasting platform, from the machine-learning and feature-engineering pipeline through to the Flask API and React dashboard used to explore the resulting forecasts.
04 / CLOSE-UP
THE ENGINEERING CHALLENGE
The central engineering challenge was forecasting sparse and highly variable ambulance demand at useful geographic and temporal resolutions. The system needed to predict both next-day demand and shorter six-hour windows while presenting uncertainty in a form that could support operational decisions.
05 / TRACKING SHOT
THE APPROACH
A plan emerges.
I developed Poisson XGBoost forecasting models using historical and temporal features including lagged demand, rolling averages, day-of-week patterns and time-of-day seasonality. The prediction pipeline converts model outputs into operational metrics such as expected calls, probabilities of future demand, surge probabilities, prediction intervals and confidence labels. These forecasts are exposed through a Flask API and consumed by a React dashboard.
06 / INSERT SHOTS
KEY FEATURES
The system takes shape.
07 / DIRECTOR'S NOTES
DESIGN DECISIONS
Things we decided along the way
01
Used a Poisson objective to model ambulance calls as count-based demand.
02
Combined lagged observations with rolling statistics to capture recent demand behaviour.
03
Included calendar and time-of-day features to represent recurring temporal patterns.
04
Produced probabilistic metrics alongside point forecasts instead of presenting predictions as certain outcomes.
05
Supported both next-day and six-hour forecasting to provide different levels of operational planning.
06
Separated the Python forecasting pipeline from the React interface through a Flask API.
design decisions
somewhere mid-build
08 / RETAKES
WHAT WENT WRONG
Naturally, not everything cooperates.
09 / FINAL SHOT
THE OUTCOME
ProblemBuildOutcome
01
Built an end-to-end machine-learning forecasting platform spanning data processing, modelling, API development and frontend visualisation.
02
Produced next-day and six-hour ambulance demand forecasts across offices and districts.
03
Extended point predictions with probabilities, surge indicators, prediction intervals and confidence labels.
04
Created an interactive dashboard through which users can upload historical data and inspect forecasts by location and time.
05
Turned historical ambulance call data into a decision-support tool for operational planning.
10 / PRODUCTION NOTES
TECH STACK
The tools behind the scenes.
PythonXGBoostPandasFlaskReactViteTailwind CSS
11 / BEHIND THE SCENES
LINKS
FADE OUT.
USER closes the file.
One project down. A few more stories left.
DIVAKAR DESSAI
Pick another?