There's a particular kind of chaos that descends when you hand an AI coding assistant a 50-million-line enterprise codebase and ask it to make a seemingly simple change. The agent might generate beautiful code—syntactically perfect, even elegant—but it lacks something crucial: the institutional memory of where things live, how they connect, and what will break three layers deep if you touch the wrong line.
Potpie, a San Francisco startup, thinks it has spotted the missing piece. On February 23, 2026, the company announced it had raised $2.2 million in pre-seed funding to build what co-founders Aditi Kothari and Dhiren Mathur describe as a "foundational context layer" for AI coding agents. Led by Emergent Ventures, with All In Capital, DeVC, and Point One Capital joining in, the round will fund early enterprise deployments and expand the engineering team—though not, perhaps, as quickly as the founders might like.
The problem they're chasing is deceptively straightforward. Modern AI can write code. It just can't read the room.
When Smart Tools Act Dumb
Kothari, now CEO, and Mathur, the CTO, started working on this in October 2023. By January 2025, they'd gone public with an approach centered on something that sounds almost academic: a knowledge graph, built in Neo4j, that attempts to unify code, tickets, logs, documentation, and code reviews into a single queryable layer. Think of it as giving an AI agent the kind of contextual awareness a senior engineer develops after years of fixing bugs at 2 a.m. and learning which database calls always timeout on Fridays.
The pitch is that without this layer, AI agents are flying blind. They can generate impressive snippets in isolation, but drop them into a real enterprise environment—with its legacy dependencies, undocumented workarounds, and that one service nobody wants to touch because the original author left in 2019—and things get messy fast.
"The real challenge is not just generating code but understanding systems deeply enough to change them safely," Anupam Rastogi of Emergent Ventures said in the announcement. He pointed to what he called Potpie's "ontology-first architecture" as the differentiator, though whether enterprise buyers will care about ontologies or just want something that works remains an open question.
What's Under the Hood
The platform itself includes a suite of prebuilt agents: debugging, code generation, Q&A, spec-driven development, integration testing. There's also custom agent orchestration for teams that want to build their own workflows. Integrations run through the usual suspects—Slack, GitHub, Jira, Sentry—and the company offers both self-hosted and on-premise deployment options, a nod to the regulated industries they're targeting.
Potpie released version 1.0.0 of its open-source core the same day as the funding announcement. The GitHub repository has pulled in 5.3k stars, running on a stack of Python, FastAPI, Celery, Redis, and Neo4j. That level of early traction suggests at least some developer curiosity, though stars don't always translate to sustained adoption.
The company is aiming at engineering organizations wrestling with codebases between one million and north of 100 million lines—Fortune 500s and publicly listed firms in healthcare and insurtech, according to their materials. Specific customer names haven't been disclosed, which is typical for early-stage enterprise software but makes it harder to gauge real-world validation.
Context as Commodity?
Potpie's timing aligns with a broader industry conversation. The New Stack observed in February that "context is AI coding's real bottleneck," and recent academic research has circled similar themes around codified context for software engineering agents. If context really is the constraint, then half a dozen well-funded startups are probably building variations on the same solution right now.
The team currently sits at around a dozen employees, according to their LinkedIn company page. A backend engineering role recently posted in Bengaluru signals they're building out capacity, though hiring in a pre-seed startup always means balancing ambition against runway.
Whether a knowledge-graph approach can solve enterprise AI's context problem—or whether this is just one more layer in an already complex stack—is the kind of question that only gets answered in production. Potpie has assembled early backing and an open-source foundation to test the thesis. Now comes the harder part: proving it works when the code is messy, the deadlines are tight, and the stakes are real.
For now, the founders are betting that the gap between what AI agents can generate and what they can safely deploy is wide enough to build a company around. Time will tell if enterprises agree.
