A lost laptop is a small emergency with a long memory. Somewhere inside a large company, an employee reports it. A support specialist knows that executives get a faster replacement, that airport losses require a particular form, and that one country has a rule the official guide forgot. The specialist fixes the problem, leaves a few words in a ticket and moves on. The employee gets a computer. The handbook remains serenely incorrect.
Eugen Alpeza has built a company around that discrepancy. Edra, the New York enterprise AI startup he co-founded with Yannis Karamanlakis, studies the residue of work: support tickets, messages, logs, emails and the unglamorous comments where judgment gets recorded. From that material, it reconstructs a set of instructions detailed enough for an AI agent to use and clear enough for an employee to inspect.
The pitch sounds practical because it began in practice. Alpeza spent seven years at Palantir, close to the difficult business of putting software inside large organizations. The lesson he carried out was not that models were unintelligent. It was that companies were insufficiently explained. A general model arrives knowing plenty about language and nothing about who approves a particular refund at 4:45 on a Friday.
“Enterprise context is the new source code.”Eugen Alpeza
That sentence is Alpeza's compact theory of the next enterprise software layer. The instructions surrounding an agent reveal how a company decides, escalates and makes exceptions. They are operational intellectual property. If those instructions live inside an opaque system, the business may be renting its own memory. If they remain readable and editable, a person can find the wrong line and repair it.
A career spent where the diagram ends
Alpeza's route into enterprise sales began on technical ground. He studied computer science at the University of Edinburgh, then computational statistics and machine learning at University College London. Before Palantir, he worked as a machine learning engineer at Skimlinks. The sequence matters: his commercial instincts were trained alongside an engineer's impatience with abstractions that fail under load.
At Palantir, Alpeza helped establish the U.S. commercial go-to-market motion and started the company's work with AT&T. The assignment put him near a sprawling organization, where a product is tested less by the elegance of its demo than by the number of local realities it can survive. He later took on the 2023 launch of Palantir's AI Platform under CEO Alex Karp.
He and Karamanlakis also created Palantir's Forward Deployed AI Engineer role. A forward-deployed engineer sits near the customer, translating research into production and learning what a business means when it says a process is “standard.” The job is part engineering, part anthropology and part tactful disbelief. It rewards the person willing to ask why the diagram has four boxes while the actual work has 40 decisions.
The forward-deployed model works because humans learn context quickly. They interview experts, watch decisions and build the software around what they discover. It is also laborious. Every deployment begins another apprenticeship. Every change in policy threatens to make yesterday's careful translation obsolete. Alpeza and Karamanlakis began to wonder whether software could participate in the learning, rather than waiting for a fresh squad of people to perform it by hand.
A 13-year conversation finds its company
The partnership predates the product. Alpeza and Karamanlakis met while studying at the University of Edinburgh and became close friends. By the time Edra announced its funding in 2026, they had known each other for roughly 13 years. They had long intended to start a company together, then spent enough time inside the same industry to discover what that company should do.
Their roles are complementary. Alpeza is CEO, carrying the commercial story and the buyer relationship. Karamanlakis, Edra's CTO, became Palantir's first Forward Deployed AI Engineer and led work that moved language models from demonstrations into production. Sequoia partner Luciana Lixandru has described Alpeza as a seller who can earn the trust of skeptical buyers, while praising the technical solidity Karamanlakis brings to the pair.
Trust between founders is easy to print on a culture page and harder to manufacture in a crisis. Their version has years of evidence behind it: university, a demanding employer, the creation of a new job function and, finally, a startup. The friendship is not decorative lore. It is part of Edra's operating system.
The manual writes itself, with an editor nearby
Edra begins where work already leaves a trace. It connects to systems such as ServiceNow, Zendesk, Jira and Outlook, then analyzes historical activity and existing documentation. Recurring resolutions become candidate instructions. Contradictions become questions. Missing subjects become visible gaps. The result is what Edra calls a library of executable knowledge.
Calling it a library is revealing. A library has shelves, labels and a way to correct the catalog. It does not dissolve the company's rules into model weights and ask everyone to have faith. An expert can read the instructions, inspect why Edra proposed them, approve a change or reject it. When new work contradicts an old rule, the system can bring the disagreement back to people.
That arrangement gives humans a better role than ceremonial oversight. They become teachers and editors. Their judgment remains visible at the point where an agent needs it. Alpeza's public shorthand is clean: “Humans teach. Agents execute.” The line also admits that neither side is finished. The employee keeps encountering edge cases; the agent keeps needing a current account of them.
“The instructions you give your agents about how your business actually runs are your IP.”Eugen Alpeza
Show us the mess for one week
In March 2026, Edra emerged from stealth and announced $30 million in funding. A $6.5 million seed round had been led by 8VC and A*. Sequoia led the later $23.8 million Series A, with other participants including HubSpot Ventures. The company named ASOS, HubSpot and Cushman & Wakefield among its customers.
Edra's public challenge: provide a sample of operational data, then inspect the processes it reconstructs.
The Series A led by Sequoia, following a $6.5 million seed round.
The funding is less informative than Alpeza's offer to prospective customers: give Edra a cut of your data, and the company will show you your processes in a week. It is an enterprise sales pitch shaped like a falsifiable claim. The customer does not have to admire a generic demo. It can look at the reconstructed instructions and say, with precision, where the software understands the business and where it does not.
HubSpot's own account supplies a view of the wedge. Its customer support organization connected Edra to agent logs and escalation data. The software identified recurring patterns and knowledge that had not reached the documentation, then suggested edits and new articles for review. This is an intentionally narrow beginning: support and IT service work create abundant records, repeat often and punish stale instructions.
The narrowness may be a virtue. Alpeza's broader ambition is a knowledge layer for enterprise agents, but a platform earns that future one useful process at a time. Lost laptops and support escalations lack the glamour of a general digital worker. They have budgets, owners, histories and an answer key. A system can prove itself there.
The argument hiding inside the product
Edra carries a view about where power will collect in enterprise AI. Models can be exchanged. The state surrounding them - institutional memory, approved instructions, past corrections and the logic of exceptions - becomes harder to move. Alpeza argues that companies should own this context in a white-box system they can read, modify and govern.
It is a sober counterpoint to the fantasy of the autonomous agent. Autonomy without context is merely confidence on unfamiliar premises. Useful agency requires a current map of the institution: which policies are real, where judgment is allowed, and how yesterday's exception changed tomorrow's procedure. An agent operating from a stale wiki is the office equivalent of a tourist navigating by a restaurant menu.
The idea also changes documentation from a periodic chore into a by-product of operations. People do not stop work to narrate everything they know. Their decisions already create evidence. The system's job is to synthesize that evidence without confusing storage for understanding. Alpeza has been explicit that keeping everything is noise; the challenge is identifying what will help with a future task.
This is where Edra's claims will be tested. Businesses contradict themselves. One expert's shortcut may be another team's prohibited habit. Old tickets can preserve bad policy as faithfully as good judgment. The white-box design matters because it provides somewhere for disagreement to land. The value does not come from pretending ambiguity vanished. It comes from making ambiguity legible enough to resolve.
New York urgency, London memory
Edra is growing in New York and London, a split Alpeza chose deliberately. The London office began with engineers he and Karamanlakis had known and worked with for years. In one public update, he said the team grew from six people to 18 in six months. He also made a practical case for the geography: East Coast and London hours overlap enough for close collaboration, while each office can retain an in-person culture.
The pattern resembles the product. Start from accumulated trust. Keep the important knowledge close. Make the handoff explicit. A young company is itself a collection of unwritten rules, and Alpeza is hiring while arguing that unwritten rules are an expensive dependency. There is a pleasing pressure in that contradiction. Edra must document the kind of company it is becoming while building software that promises to help customers do the same.
Alpeza's aspiration is not modest in scope, but it is specific in mechanism. He wants Edra to become the learning layer that keeps enterprise agents aligned with the way a company operates now. The wager rests on an ordinary observation: every day, employees teach the business what it is. They do it in replies, corrections, approvals and workarounds. The knowledge is already arriving. It simply needs a place to become durable.
The lost laptop will still be lost. Someone will still know the peculiar rule. If Edra works as Alpeza intends, the next person - human or agent - will not have to rediscover it from the floor.