The Company That Learned to Read the Old Code

Somewhere in an old application, a programmer decided that an order above a certain amount needed a human's approval. The rule made sense to somebody. Twenty years later, the programmer is gone, the documentation is missing, and a new team is ready to replace the software. Their AI assistant can write the new approval screen in seconds. It cannot know which orders the old system quietly held back unless somebody finds that rule first.

That small, dull act of discovery is the business Crowdbotics has grown into. The Berkeley company began by helping customers produce applications more quickly. It assembled reusable code, wrote product requirements with AI and offered people to manage the build. In September 2025, it introduced a new name, CoreStory, for a different emphasis: read the code already in a company, recover its logic, and make that knowledge available to engineers and coding agents.

The first customer problem was speed

Crowdbotics was founded in 2016, according to its company profile. Its founder and CEO, Anand Kulkarni, had previously built LeadGenius, an AI-assisted sales data company. With Crowdbotics he took aim at a familiar annoyance: teams repeatedly write software that already exists in some form. A login flow, a payment connection, a scheduling page - none is effortless, but few need to be invented from first principles every time.

The early pitch was systematic code reuse. Product managers could describe an app, the platform could help produce a product requirements document, and developers could assemble from a catalog of modules while retaining access to full source code. Managed development services added project management and engineering help. That combination sat between an agency and a software platform: a customer was buying a faster route to a working app, not merely a drag-and-drop mock-up.

It found serious work. A federal record shows a $749,601 Air Force SBIR Phase II award in 2021 for machine-learning analysis of pilot maneuvers and training. This was not an abstract promise to “transform defense.” It was a funded attempt to make assessment of flight performance less dependent on subjective review. Crowdbotics later reported more than 500 customers, with the Air Force among them, at the time of its January 2023 funding announcement. That historical count is useful as a measure of the earlier business, not a present-day CoreStory customer tally.

Then the bottleneck moved

By 2023, Crowdbotics had raised a $40 million Series B led by NEA to expand its enterprise business. Its CodeOps story was still about making new software faster: write requirements, match them to existing code and deploy a full-code application. A 2024 collaboration with Microsoft brought that workflow toward Azure and GitHub. The company argued that requirements were the real hard part of development, because a developer cannot efficiently build a thing the organization has not agreed to describe.

That observation had a second life. Enterprises do not start every project with a clean requirements document. They inherit databases, message queues, half-remembered exceptions and millions of lines of code. The specifications are often inside the application itself. When an organization asks an AI agent to modernize such a system, the agent's fluency can be a liability: it can confidently produce a plausible replacement for behavior it never understood.

“Crowdbotics was about building apps. CoreStory is about understanding code.”CoreStory's 2025 rebrand announcement

The new name made that change unusually explicit. CoreStory's product scans source code, builds a persistent model of architecture, workflows, dependencies and business rules, and translates those findings into readable requirements. Developers can ask questions of the model; product managers can use the resulting specifications; coding assistants can query it through integrations. The company lists COBOL, Java, C#, C++, JavaScript, PHP, Python and other languages among its targets. It is trying to be the memory of a system before it becomes the brief for a new one.

A pet store with 95 secrets

The company's most revealing public demonstration uses an unglamorous artifact: Sun Microsystems' Java Pet Store 1.3.2, a J2EE reference application from 2003. It is small enough for outsiders to inspect, with about 276 classes, but old enough to contain the architectural habits of its period: message-driven beans, JMS and XML deployment descriptors. CoreStory ran its modernization playbook against it and reported 95 business rules across ten domains. Ten were marked as critical migration risks.

One example concerns locale-specific auto-approval thresholds tucked into a method. Another concerns an order lifecycle spread across six message-driven components. A rewrite that preserves the screen but loses either rule has failed in a way a pleasant demo might miss. CoreStory also found no circular dependencies and no shared database tables in the reference app, facts that made separation into services more feasible. The practical point is less the score the company assigned than the questions it asked: which behavior is real, where does it live, and what would break if it vanished?

The demonstration is also candid about its limits. Java Pet Store is a reference application, not a sprawling bank or airline estate. An experienced architect could inspect a system that size by hand. CoreStory's claim is that the same sort of inspection becomes much harder at enterprise scale, and that a persistent model keeps an answer useful after the first meeting. In the demo, a separate coding tool wrote the migration code; humans made architectural calls. That division of labor is central to its pitch.

Where the bet works

CoreStory sells to organizations with software too important to guess about: financial institutions, healthcare operators, manufacturers, government teams and other enterprises maintaining long-lived applications. The buyer might be an architect planning a migration, a developer tracing a defect, a product manager reconstructing requirements, or a team onboarding engineers to a codebase its original authors have left. Its alternative is familiar: weeks of manual code archaeology, stale wikis, expensive consulting, or a general AI assistant supplied with fragments of a repository.

The distinction matters most when the code has consequences. For a new landing page, an AI assistant with a good prompt may be enough. For a regulated system with obscure exceptions, the source code needs to be treated as evidence. A useful pilot would pick one consequential workflow, ask CoreStory to identify its rules and dependencies, then have engineers check the findings against the source and tests. The output should tell the team what to preserve, where that behavior lives, and which uncertainties still require a human decision.

That is a slower-sounding promise than “build an app in minutes.” It may also be the more durable one. The market is crowded with tools that make writing code easier. Crowdbotics' interesting turn is to argue that the next expensive mistake happens before the first new line is written - when a team mistakes a readable program for an understood one.

Anand KulkarniFounded Crowdbotics and leads its CoreStory chapter. The company's public writing repeatedly returns to the same premise: knowing what to build is harder than typing the code.