Four Weeks to a Single Customer View: How a Lakehouse Architecture Solved an Old Problem for Advisors

TL;DR

A small AppDirect engineering team built a unified view of customers, orders, and quotes across three business units in four weeks, with no manual coding. Advisors now have one place to see every customer across hardware, software, and connectivity. It runs on a Lakehouse architecture, and a person reviews every duplicate match before any records merge. A project like this would normally be scoped in quarters.

An old problem, finally solved

AppDirect advisors serve customers who buy hardware, software, and connectivity, but the systems behind those product lines don't talk to each other. A single customer can look like three different customers depending on which screen an advisor is on. Cross-sell opportunities hide in plain sight. Basic questions go unanswered.

A small AppDirect engineering team fixed that in four weeks, on a data foundation called a Lakehouse architecture. It's a concrete example of what AppDirect’s SVP of Engineering Mathew Spolin recently described in How to Adopt AI Development at Scale: an engineering organization that went from 0% to over 90% AI-generated code in a single year, with a 44x increase in accepted lines and quality metrics holding or improving as velocity climbed.

Those numbers invite a fair question: what does 90% AI-generated code actually build? This project is one answer, built for the advisors who use our platform every day to recommend technology to their customers.

Siloed data hid cross-sell opportunities

“An advisor selling hardware, software, and connectivity had no single place to see a customer. You couldn't answer basic questions. How many unique customers do I have? Does this one buy from more than one product line? What do they spend with us in total?” —Mathew Spolin, SVP of Engineering, AppDirect

Advisors recommend technology to solve their customers' problems. A good recommendation depends on knowing who each customer is and what they already buy.

“An advisor selling hardware, software, and connectivity had no single place to see a customer,” said Mathew Spolin, SVP of Engineering at AppDirect. “Accounts, orders, and quotes lived in separate systems. You couldn't answer basic questions. How many unique customers do I have? Does this one buy from more than one product line? What do they spend with us in total?”

Cross-sell opportunities stayed invisible unless someone dug for them by hand. Company-wide customer metrics were unreliable because the underlying records disagreed.

Most companies live with this kind of problem for years. It touches enough systems that a traditional rebuild gets scoped in quarters.

What we built: A unified view, end-to-end

The team delivered a working end-to-end slice across the three categories advisors sell: connectivity, cloud software, and hardware lifecycle management.

Behind that single view sits a complete data pipeline that:

• Ingests raw customer, order, and quote data from each underlying system

• Cleans and deduplicates records using probabilistic matching.

• Routes every likely match to a human review queue, where a reviewer confirms or rejects it

• Produces a "golden record" for every customer, the authoritative version the rest of the business can trust

• Surfaces all of it through a consolidated advisor interface

“Advisors now get one view of their customers, quotes, and orders across product categories, instead of a separate tool for each,” Spolin said.

The stack, in plain English

The platform sits on top of a Lakehouse architecture, a modern data approach that combines a data lake (cheap, flexible storage for raw data of any shape) with a data warehouse (structured, query-ready tables for analytics). It’s akin to a single library that accepts both loose notes and bound books, and lets you search across all of it. AppDirect's Lakehouse runs on Databricks, built out by our data platform team as a foundation other projects can now build on.

Inside the Lakehouse, data flows through a Medallion architecture, a three-tier convention named after Olympic medals:

• Bronze: raw data, exactly as it arrived from the source system

• Silver: cleaned, deduplicated, standardized

• Gold: the polished, trustworthy "single source of truth" used by the business

For the hard problem of figuring out whether "Acme Corp," "ACME Corporation," and "acme corp." are the same customer, the team used Splink, an open-source probabilistic matching library that assigns a confidence score to potential duplicates rather than making a binary yes/no call.

The advisor-facing experience was built using a micro-frontend approach: small, composable UI pieces that let the team ship a consolidated view without rebuilding the surrounding tools.

None of these are exotic choices. The team put them together in four weeks.

Why humans stay in the loop

Nothing merges automatically.

When Splink flags two records as likely the same customer, the match goes to a human review queue. “A reviewer looks at each match and picks which values win where the records disagree,” Spolin said. “The original records stay in the system, so every merge can be traced back.”

The team chose this on purpose. Billing, contracts, and account history all depend on customer records, and a bad merge that nobody notices costs far more than a few minutes of review. AI wrote the code. People decide which records are the same customer.

Four weeks, two sprints, no manual coding

The team started from the original database schemas of the source systems. From there, the engineers built everything through Cursor, the AI-native IDE AppDirect standardized on in early 2025. Scope expanded iteratively as new business units were brought in. The full slice, covering provider sales, the marketplace, and Firstbase, shipped in four weeks across two sprints.

A project that touches multiple business systems, adds a new data platform layer, and ships a new advisor interface on top would usually take months just to scope. “Normally we'd spend a quarter planning something like this before writing any code,” Spolin said. “This time, advisors had a working version before that planning would have been done.”

Spolin described this shift in How to Adopt AI Development at Scale: "The cycle time from idea to merged code has compressed, and the craft is shifting toward architecture, assumptions, and product thinking."

The engineers on this project didn't spend four weeks typing code. They spent it making decisions: what to ingest, how to model it, where to put the human review boundary, how to expand scope when each new business unit came online. Cursor handled the implementation work that used to take up most of the calendar.

“Normally we'd spend a quarter planning something like this before writing any code. This time, advisors had a working version before that planning would have been done.” —Mathew Spolin

What it shows

A company-wide number like 90% AI-generated code is easy to doubt. A specific system, and the advisors it helps, is harder to argue with.

This project is a useful proof point:

• It solves a real, long-standing user problem. Advisors have lived with fragmented customer data for years. This gives them one screen, real numbers, and visible cross-sell opportunities they couldn't see before.

• It's a foundation, not a one-off. The Lakehouse and golden customer records are built for other teams to use.

• It spans three business units with different source systems, data shapes, and advisor workflows.

• The design is deliberate. People review every merge, and the original records stay auditable.

• It took weeks instead of months, the same shift the org-wide numbers show, in a single project.

For companies deciding whether AI-native engineering lives up to the claims, this is a concrete example. The tools work. The harder part is the organizational change behind them: the metrics, the habits, and how engineers spend their time.

The new view reaches advisors in October 2026.

More reading

• How to Adopt AI Development at Scale: AppDirect's Leap from 0% to Over 90% AI-generated Code—The organizational transformation behind projects like this one

• Build vs Buy Software—How AI-enabled Software Development and Vibe Coding are Changing the Game—The strategic framework for build-vs-buy in the AI era