Inspiration

Before writing a line of code, we interviewed 15 consultants at boutique and mid-sized firms. Every one of them told a version of the same story: an analyst rolls off a project, the replacement inherits a Drive folder and a Slack history, and spends the first two weeks reconstructing decisions someone already made. The files survive the handoff. The reasoning doesn't.

AI made this worse. A growing share of the thinking on a project now happens in a ChatGPT or Claude tab that nobody else on the team can see, and the AI itself has no idea what the team already decided. The tools firms are adopting to fix this index the entire company, which means an AI agent can surface a document to someone who was never cleared to see it. For firms handling confidential client data, that is not a productivity problem. It is an ethics and governance problem.

We built Kerniva to answer a narrower question: what would an AI workspace look like if it were scoped to a single project, built around the handoff, and designed so humans never lose control of what the AI sees or changes?

What it does

Kerniva is a multiplayer AI workspace for consulting teams. Each project gets one shared space where files, research, AI sessions, tasks, and decision history live together, so any team member, or the AI, can pick up where someone else left off.

Library: the project's files and research, in one place. Brain: an AI that works from the full project context. Every retrieval passes through a clearance filter before the agent sees it, so the AI can only pull documents the current user is authorized to access. Relay structured handoffs between team members that carry the reasoning, open questions, and decisions, not just the deliverables.

• Tasks: project tasks tied to the files and sessions they came from.
• History: git-style versioning of every edit, including edits the AI makes. Anyone can diff what changed, see whether a human or the AI changed it, and roll it back.
• Skills: reusable workflows for recurring deliverables.

Brain's clearance-filtered retrieval and history are the core of our Track 1 case: a firm can hand real client work to an AI agent while keeping a complete, auditable, reversible record of everything it touched, and a hard guarantee that it only ever saw what the user was cleared to see.

How we built it

• Frontend: framework
• Backend: framework with database for project data and vector store for retrieval
• AI layer: Anthropic's Claude models via the Claude for Startups program, with a retrieval pipeline that applies the user's clearance level as a filter before the query runs, not as a post-hoc check on results
• Versioning: History is modeled on Git. Every edit, human or AI, is a commit with a diff, an author, and a parent, so rollback is a pointer move rather than a restore.
• Infrastructure: Azure, provisioned with OpenTofu, supported by Microsoft for Startups and NVIDIA Inception

We built the MVP as a three-person founding team: one on the core product and dev environment, one on the website, deployment, and parts of the product, and one on customer development and growth. No outside contributors.

Challenges we ran into

Filtering before retrieval, not after. The obvious way to do access control in a RAG system is to retrieve the top results and drop the ones the user can't see. That's wrong: the model's context has already been shaped by documents the user wasn't cleared for, and the result set gets thinner for low-clearance users. We had to push the clearance filter into the retrieval query itself, which meant rethinking how we index and rank.

Versioning AI edits that aren't line-based. Git works because text has lines. AI edits to a structured deliverable (a slide, a table, a model) don't differ cleanly the same way. Getting History to produce a readable diff of what the AI changed took several iterations on the data model.

Earning trust without traction. Consulting firms are conservative about tooling, and we had no customers to point to. We got our first pilot by being candid about what the product could and couldn't do yet, and by showing the audit trail rather than describing it.

Accomplishments that we're proud of

• A working MVP with all six modules, built by founders only.
• A live pilot since September 2026 with a boutique consulting firm, on real client work, not a sandbox.
• Early conversations with a partner's team at KPMG.
• Accepted into Claude for Startups, NVIDIA Inception, and Microsoft for Startups; completed AltaLab's Fall 2026 cohort.
• A clearance-filtering design we believe is more correct than what most AI tools ship with by default.

What we learned

The ethics of enterprise AI mostly aren't about model alignment. They're about plumbing: what the model is allowed to see, whether a human can tell what it changed, and whether that change can be undone. Those are engineering decisions, and most tools get them wrong by default because the failure is invisible until something leaks.

We also learned that nobody asks for "AI handoff" by name. They ask for the two weeks back. The job is to solve that and let the governance come along quietly.

What's next for Kerniva

• Convert the current pilot into a paying customer and onboard the KPMG team.
• Expand from boutique U.S. consulting firms into adjacent professional services with the same confidentiality constraints.
• Open the clearance-filtered retrieval layer so firms can plug in their own access-control policies.

Built With

• anthropic
• azure
• claude
• cloud-foundry
• linux
• opentofu