A car dealership can change its name without changing its address. It can change its owner without changing its brand. Somewhere in a collection of old records, all three versions may continue to exist, each apparently a different business. Ask an AI system to value that dealership and you have given it a small identity crisis before it has done any arithmetic.
This was the decidedly earthly problem behind one Shorthills AI project: decades of information covering more than 18,000 dealerships, with inconsistent formats and missing identifiers. The company’s JumpIQ case study reports that a full refresh took more than a week. Its response involved matching names, normalizing addresses and deciding which conflicting record should survive. The resulting platform, JumpIQ, brought the refresh down to about eight hours.
- The business: custom enterprise AI, from data preparation to operating agents.
- The buyers: teams with complex documents, scattered records and repetitive expert work.
- The useful lesson: measure the entire workflow, including human review.
There is something pleasing about an AI story whose supporting character is an address. The glamorous part of the industry generates sentences. The commercially useful part often begins by ensuring that two rows in a database refer to the same thing.
The database gets the first speaking part
JumpIQ runs on a Databricks data foundation. Shorthills combined source records, built more than 150 standardized signals per dealership, and added valuation, forecasting and scenario tools. Analysts can compare businesses, explore markets and export diligence summaries. The reported speed improvement matters because a valuation workflow needs information it can actually refresh.
The engineering sequence is instructive. First settle identity. Then make measurements consistent. Then build the prediction and the interface. A business that skips the first two steps can produce a beautiful dashboard in which every department disagrees about what the numbers mean. Shorthills’ contribution sits across that sequence, rather than ending when a model produces an answer.
Who pays for the unglamorous work?
Shorthills sells to enterprises: legal researchers, finance teams, healthcare organizations, retailers and automotive businesses. Its homepage includes testimonials from Dave Cantin Group and British bargain community LatestDeals.co.uk, as well as a partnership statement from PwC India. These are different kinds of organizations, but they share a familiar predicament: useful information exists, and getting it into the right hands takes too much work.
“Their professionalism, consistency, and communication is excellent.”Tom Church, co-founder, LatestDeals.co.uk
The company’s services describe a fairly complete engagement. Strategy work identifies worthwhile use cases and a roadmap. Data engineers prepare the inputs. Agent developers connect models to business tools. Managed operations deal with monitoring, guardrails and costs after deployment. That last stage is particularly revealing. A demonstration has an audience; an operating system has a queue of jobs, users with permissions and a monthly bill.
For buyers, the appeal is having one provider work across these boundaries. For Shorthills, the model combines custom delivery with reusable software. An accelerator gives the next project a head start while leaving room for the awkward details of the customer’s systems. Its public offering reads like a services business accumulating products through repeated encounters with similar problems.
The lawyer needs a better search party
Consider legal research. A fluent answer is not much comfort if it omits the case that defeats your argument. In its March 2026 case study, IBM describes a Shorthills assistant built with watsonx.data and Langflow, combining keyword, vector and graph search. Queries are routed according to what they need: a precise identifier, a related meaning or a network of legal relationships.
The architecture ingests documents, extracts entities and supplies citations with retrieved results. IBM reports improvements in completeness and legal-reasoning diversity for that implementation. The broader lesson is architectural: one retrieval method does not have to carry every question. The system can also support on-premises deployment, a useful option when sensitive material must stay within an organization’s environment.
A specific document or identifier.
A question about meaning.
Connections between legal entities.
This is a good place to draw a distinction between accuracy in conversation and usefulness in professional work. A lawyer wants a trail back to the material. A business wants a deployment arrangement its security team can accept. Both requirements belong in the design before anyone starts admiring the prose.
A customer has five things to say
Shorthills’ support-chat case offers a smaller, equally revealing problem. Customers do not reliably present one tidy request per message. They may ask to add, remove and relocate items in the same breath. A bot built around a single intent can turn that perfectly ordinary sentence into a succession of clarifications.
The company describes fine-tuning an open-source model on more than 15,000 synthetic examples, allowing it to extract up to five intents in a turn. A separate rules layer matches names against a catalog exceeding 40,000 products and validates the requested actions. The output is a structured action list for the backend. Shorthills reports roughly 1.2-second latency and an August 2024 production launch after a two-month build.
The copyable detail is the rules layer. Understanding a request and authorizing its execution are different jobs. The model supplies an interpretation; catalog matching and business validation decide whether that interpretation can become an action. That division gives engineers something concrete to inspect when a customer says the bot did the wrong thing.
A consultancy with a product drawer
Among its reusable tools, KodeBricks targets Databricks development. Developers describe a task in conversational language; the accelerator helps create pipelines and orchestrate workflows from their existing development environment. Shorthills’ product page focuses on the everyday friction of setup, boilerplate and switching between tools. The interesting promise is less time spent negotiating the machinery around the code.
Databricks includes KodeBricks in a roundup alongside migration tools from Persistent Systems, LTIMindtree, Xebia and others. That places Shorthills in a competitive field of implementation specialists. The overlap concerns a practical purchasing decision: whether to build internally, hire an integrator or buy help that comes with reusable components.
Other tools move closer to specific professions. BayesPair connects family-history intake, digitization of hand-drawn pedigree charts and chart design. Label Gaze reviews packaging labels and presents findings with regulatory citations, while allowing people to edit or override them. A family tree and a package of cosmetics may seem unlikely neighbors. Both contain information that must be interpreted within a particular set of conventions.
Two founders, plenty of engineering
Paramdeep Singh and Pawan Prabhat founded Shorthills AI in 2018 after previously co-founding EduPristine, an accounting-training business. The company takes its name from Short Hills, New Jersey, and now reports more than 300 engineers across its global operation. Its founders’ biographies span technology and business domains, an appropriate combination for work that begins with software and ends with someone’s operating process.
The careers page’s open roles include data engineering, backend development and data science. Company-published employee accounts emphasize learning and collaboration. The job mix offers a useful glimpse of the work itself: AI delivery requires people who can move data, build applications and maintain infrastructure as well as people who can tune a model.
The budget includes the exceptions
A buyer evaluating this kind of engagement should cost the whole job: source-system access, data cleanup, integration, computing, security review and ongoing operation. The dealership example measures processing time. The support example measures response time. Neither measure, by itself, is a calculation of total financial return. A faster workflow creates value when the saved capacity can be used.
Shorthills’ transaction-analysis work illustrates why human capacity belongs in that budget. Its system ingests ledger and invoice data, applies classification and tax logic, and sends exceptions to reviewers. Experts correct outputs and feed those corrections back into the process. The target is a manageable review queue. A system that automates ordinary cases while flooding people with uncertain ones may simply relocate the bottleneck.
The company’s governance service covers monitoring, access controls, evaluation and cost oversight. These are recurring responsibilities. Models can change, business rules can change, and incoming data can acquire new surprises. Implementation therefore needs an owner after launch, with the authority and time to respond.
For a prospective customer, a useful pilot starts with one bounded workflow and a recorded baseline. Count minutes, errors, exceptions and human touches. Make source records traceable. Decide which actions need approval. Then compare the finished process with the original one. This is an inference from the projects, and it travels well beyond this particular vendor.
The approach suits organizations that can provide accessible data, domain expertise and reviewers. KodeBricks also depends on a Databricks-oriented environment. A loosely defined task with inaccessible records offers considerably less to work with. Shorthills’ most persuasive examples begin with a named obstruction and end with a measurable change. Sometimes the obstruction is simply a business that appears under three names. Somebody still has to notice.