Databricks
Python (Django)
React
AWS S3
Gemini
Tech Stack
Client Profile
Industry
Automotive
Region
North America
Technology
Databricks
Overview
A leading automotive advisory firm that provides M&A and investment insights for the U.S. car dealership market struggled to leverage its raw data, coming from over 18,000 dealerships spanning decades. Each record had roughly 150 fields drawn from Polk, Helix, demographic and population datasets and other open sources and APIs. This had issues of inconsistent formats, missing common identifiers that prevented easy merging, and large gaps. These problems slowed extraction of actionable insights: full data refreshes took more than a week and blocked timely, strategic decisions such as dealership valuations.
To resolve the client's data challenges, Shorthills AI developed JumpIQ, an AI-powered platform that ingests and processes raw data from Polk, Helix, and other open APIs directly into Databricks. A robust data engineering pipeline was built for intelligent merging (using techniques like fuzzy matching and address normalization), cleaning, mapping, and formatting to create a unified “golden record” for each dealership. On this refined data foundation, advanced AI/ML models were deployed for predictive analytics, including revenue forecasting, sales efficiency, dealership valuation, and performance scoring—all accessible through a web-based dashboard offering detailed analytical reports and visual insights.
As a result, the client reduced data processing time from over a week to just 8 hours, gained a single clean and accurate database, and obtained significantly stronger predictive insights that enable faster, more confident strategic decisions.
Real-Time M&A Intelligence for 18,000+ Dealerships
Our Solutions
Data Foundation: Lakehouse & Entity Resolution
We stood up a Databricks-powered lakehouse with medallion layers (bronze → silver → gold) and survivorship rules to reconcile conflicts. Fuzzy matching plus brand/state heuristics created a durable golden dealer record across renames, mergers, and closures—an analytics-ready backbone with end-to-end lineage.
Signals & Feature Engineering
On unified records, we built a reusable catalog of 150+ signals per dealership spanning performance, market, and macro indicators. Features are standardized across brands/states and versioned over time, so valuations, forecasts, and benchmarks stay fair and reproducible.
Valuation & Forecasting Engines
A model suite blends store performance with market signals to produce explainable valuations and forward-looking forecasts. Scenario/sensitivity views test brand, geography, and macro assumptions—accelerating buy/no-buy calls with consistent methodology.
Delivery Experience: Analyst App for M&A Workflows
A secure analytics app streamlines real M&A tasks: search/filter/compare, geospatial views, and exportable diligence summaries. Built on governed tables and shared definitions, it keeps every stakeholder aligned—from board decks to deep dives.
Modernizing consumer support chat with a multi-intent LLM—parsing up to 5 intents with ~1.2s latency to lift automation and first-contact resolution.
Industry
Customer Experience
Region
North America
Technology
AWS
Executive Summary
A high-volume consumer support operation needed a chatbot that could accurately understand complex, multi-intent requests and respond near-instantly. The legacy bot misread intents, handled only one action at a time, and responded slowly—pushing users to call centers. We built a custom multi-intent LLM engine exposed via API that parses up to 5 intents in one message, fuzzy-matches products from a 40k+ catalog, and returns clean JSON for backend execution. Optimized with vLLM and FastAPI, latency dropped to ~1.2s, automation rose, and first-contact resolution improved—cutting support costs while lifting customer satisfaction.
Tech Stack
GPT 4
Llama 3
vLLM
FastAPI
Docker
Python
AWS
My SQL
Executive Summary
A high-volume consumer support operation needed a chatbot that could accurately understand complex, multi-intent requests and respond near-instantly. The legacy bot misread intents, handled only one action at a time, and responded slowly—pushing users to call centers. We built a custom multi-intent LLM engine exposed via API that parses up to 5 intents in one message, fuzzy-matches products from a 40k+ catalog, and returns clean JSON for backend execution. Optimized with vLLM and FastAPI, latency dropped to ~1.2s, automation rose, and first-contact resolution improved—cutting support costs while lifting customer satisfaction.
Tech Stack
GPT 4
Llama 3
vLLM
FastAPI
Docker
Python
AWS
My SQL
Modernizing Leading U.S. Automotive M&A with Databricks—unifying data from 18,000+ dealerships into golden records to deliver explainable valuations, standardized forecasts, and 8-hour refreshes
Industry
Automotive
Region
North America
Technology
Databricks
Databricks
Python (Django)
React
AWS S3
Gemini
Tech Stack
Executive Summary
A leading U.S. automotive advisory firm struggled to turn decades of raw data from 18,000+ dealerships—spread across Polk, Helix, demographic datasets, and multiple APIs—into actionable insights. The fragmented and inconsistent data made full refreshes take over a week, delaying critical decisions like dealership valuations. Shorthills AI developed JumpIQ, an AI-powered platform that ingests this data into Databricks, creating unified “golden records” through intelligent cleaning, mapping, and merging. Advanced AI/ML models then deliver predictive analytics via a web dashboard with detailed reports and visual insights. The result: data processing dropped from over a week to 8 hours, the client gained a single accurate database, and predictive insights now support faster, more confident decisions.
Modernizing Leading U.S. Automotive M&A with Databricks—unifying data from 18,000+ dealerships into golden records to deliver explainable valuations, standardized forecasts, and 8-hour refreshes
Industry
Automotive
Region
North America
Technology
Databricks
Tech Stack
Databricks | Python (Django) | React | AWS S3 | Gemini
Executive Summary
A leading U.S. automotive advisory firm struggled to turn decades of raw data from 18,000+ dealerships—spread across Polk, Helix, demographic datasets, and multiple APIs—into actionable insights. The fragmented and inconsistent data made full refreshes take over a week, delaying critical decisions like dealership valuations. Shorthills AI developed JumpIQ, an AI-powered platform that ingests this data into Databricks, creating unified “golden records” through intelligent cleaning, mapping, and merging. Advanced AI/ML models then deliver predictive analytics via a web dashboard with detailed reports and visual insights. The result: data processing dropped from over a week to 8 hours, the client gained a single accurate database, and predictive insights now support faster, more confident decisions.
Outcomes
Unify all your disparate sources into a governed data lakehouse, resolve duplicates to a single “golden record,” and standardize key signals so analysts can trust the data. That’s how we built JumpIQ for a leading U.S. automotive M&A firm: we consolidated decades of data across 18,000+ dealerships, cut refresh time from 7+ days to ~8 hours, and engineered 150+ metrics per store. On top, we added explainable valuation and forecasting models so you can run what-ifs on brand, geography, and macro factors. The result: faster, defensible diligence with scenario planning directly from your historical data.
Drastic Speed Improvement
Full data ingestion and refresh cycles reduced from over a week to 8 hours.
Enhanced Predictive Accuracy
Unified, clean database for 18,000+ dealerships, each with ~150 data points.
Comprehensive & Accurate Data
More reliable forecasts for Key Performance Indicators, sales, and valuations.
Challenges
Consumer support operations see complex, multi-step requests that legacy bots misread, handle one-by-one, and answer slowly—driving escalations and cost. Fragmented intent handling and weak product matching stall self-service and hurt CSAT. With fast, multi-intent understanding and catalog-aware actions, teams resolve more in a single turn and keep customers in channel.
Inaccurate intent recognition
Conversations stalled or escalated due to frequent misclassification.
No multi-intent handling
Users had to start separate flows for add/remove/relocate actions in one request.
High latency & operational cost
Slow replies and frequent hand-offs to agents increased workload and spend.
Outcomes
A high-volume support team was losing customers to slow, one-track chats that misread intent and pushed calls to agents. With Shorthills AI’s multi-intent engine, one message can cover add/remove/relocate and more—parsed correctly and executed together. Latency falls to about 1.2 seconds, keeping conversations fluid and reducing abandon. Automation and first-contact resolution rise as fewer queries need hand-offs, cutting support costs. Fuzzy catalog matching ensures the right products and packs are actioned, reducing rework. The net effect: faster replies, fewer escalations, and higher customer satisfaction—delivered via an API the team can extend as new products launch.
Near-instant responses
latency reduced to ~1.2 seconds for fluid chat UX.
Higher automation & FCR
more issues resolved in one interaction via multi-intent parsing.
Lower support cost
fewer escalations to human agents and easier updates for new products/services.
Frequently Asked Questions
What Shorthills AI Did
We replaced the brittle bot with a single, fast brain for chat. It understands up to five intents in one message, matches product names against a 40k+ catalog, and returns a clean, ready-to-run JSON action list. Exposed as a simple API, it drops into existing flows, and is tuned for low latency—so customers get near-instant answers and agents see fewer escalations.
We fine-tuned an open-source model on 15k+ synthetic examples to extract multiple intents and entities from a single query—scaling to 5 intents per turn.
Engineered a Multi-Intent LLM Core
We added a rules layer that fuzzy-matches extracted names to a 40k+ product catalog, validates business logic, and formats a clean JSON payload for downstream execution.
Built Precision Post-Processing & Fuzzy Matching
We packaged the model with vLLM and FastAPI, containerized with Docker, and exposed a simple API that partners can drop into existing workflows.
Delivered a High-Throughput Inference API
We applied prompt engineering, input cleaning, and response validation to consistently hit ~1.2s end-to-end latency under load. Go-live followed a 2-month build; production since Aug 2024.
Optimized for Reliability & Speed
Also Read
.jpg)
