Nilou Salehi

Salehi is Across AI's co-founder and Chief Product Officer. The company sells "agentic memory" to enterprise sales teams - software that reasons over a company's own data, holds onto what matters, and quietly drops what's gone stale. The pitch is aimed at chief revenue officers and the reps under them who burn hours updating pipelines and researching accounts. Across wants those hours back.

What makes the venture worth a second look is the resume behind it. Salehi isn't a growth-hacker who discovered AI when it got hot. She is an associate professor at the University of California, Berkeley's School of Information, affiliated with the Berkeley AI Research Lab and the EECS department, and a published authority on human-computer interaction. She studies the seam where people and algorithms meet, and whether the algorithms can be trusted there.

To build Across, she did something professors rarely do: she paused the tenured career. She is on leave from Berkeley, leading a product team of ex-Google people, UX researchers, and designers. The lecture on human-centered AI became the spec sheet.

The co-founders complete the triangle. Steven Mih, the CEO, carries 18 years of quota-carrying sales experience and once sold his startup Ahana Cloud to IBM. Afshin Nikzad, the CTO, holds a dual Stanford PhD in algorithms and machine learning and is, by reputation, a competitive-programming gold medalist. Salehi is the bridge between the salesperson's instinct and the algorithm's logic.

Long before agentic memory, there was Dynamo. As a doctoral researcher at Stanford, Salehi helped build an organizing platform for the invisible workforce behind Amazon Mechanical Turk - the people who label data and answer surveys for cents. Her 2015 paper "We Are Dynamo" studied how to spark collective action among workers who never meet. It is a strange thing for a computer scientist to study, and that is the point.

She also built Hive during her PhD: a system that sorts a community into small teams and then rotates the membership using an optimization algorithm, so viewpoints keep mixing instead of hardening. Mozilla picked it up to improve Firefox accessibility. The throughline of her work is not the code. It is the question of how software changes the way humans cooperate.

Her path to that question started in Iran, where she studied computer engineering at Sharif University of Technology, one of the country's most competitive programs. From there to Stanford for the PhD she finished in 2018, then straight to Berkeley as faculty.

At Berkeley the research kept widening: equity in San Francisco's school-assignment algorithms (work that made her a W. T. Grant Foundation Scholar), restorative-justice approaches to online harm, anti-Muslim hate speech, and reliable machine translation for high-stakes settings, the last in collaboration with Timnit Gebru. The translation work won an Outstanding Paper Award at EMNLP 2023 and asked exactly the question Across now sells an answer to: when should a person trust what the model just told them?

That medical-translation project is the clearest tell. Her team built quality-estimation models and interface designs meant to help a professional spot the moment an AI output should not be trusted - the moment to slow down and check. It is a deeply unglamorous problem. A mistranslation in a hospital is not a benchmark score; it is a person. The lesson she carried out of that work is that the model's raw accuracy is only half the system. The other half is whether the human on the other end can see what the machine did and decide whether to believe it.

The standard way to feed a company's data into a model is retrieval-augmented generation, or RAG: when you ask a question, the system fetches relevant documents and stuffs them into the prompt. It works, until the fetched document is six months out of date and the model answers with total confidence anyway. In a sales context, that is the difference between a warm follow-up and an embarrassing one.

Across positions its agentic memory as the next step past that. The system reasons over a company's own enterprise data, continuously adapting what it holds: retaining the information that still matters to a deal and discarding what has gone stale. The promise is contextually accurate insight that keeps pace with a live pipeline rather than a snapshot of one.

The buyer is the enterprise sales organization - chief revenue officers and the account executives who report to them. The complaint Across is built around is a familiar one. Steven Mih, the CEO, frames it from experience: deep product and account knowledge always gave him an edge in prospect conversations, but assembling that knowledge took an enormous amount of time. Across is the attempt to compress that time without compressing the accuracy.

This is where Salehi's academic spine becomes a product advantage rather than a footnote. A founder who has spent years studying when humans should trust algorithms is exactly the person you want designing a tool whose entire value rests on a salesperson trusting it in front of a customer. Her insistence on control and transparency - that a user must be able to see what an agent did and steer what it does next - is not a compliance checkbox bolted on at the end. It is the design premise.

There is a tension in what Salehi is attempting, and she does not pretend otherwise. Academic HCI moves at the speed of peer review. A seed-stage startup moves at the speed of payroll. She has chosen to drag the patient discipline into the impatient arena and see if the values survive the transfer.

Her stated ambition is unusually specific for a founder. She does not talk first about market size or revenue. She talks about leaving a mark on the industry defined by trust, reliability, and transparency - about being at the forefront of a human-centered AI ethos. In a field where "agentic" too often means "autonomous and unaccountable," she is arguing for agents that show their work.

The bet underneath Across is that this is not a constraint on the business but the business itself. Capability is becoming a commodity; every vendor has access to similar models. What is scarce is the boring, hard-won trust that lets a professional act on what the machine says. Salehi spent a decade studying that scarcity. Now she is selling the cure.

It is early. Across came out of stealth at the end of 2024 and is hiring engineers across San Francisco and Vancouver. The thesis is unproven at scale, the market is crowded, and the leave from Berkeley is, for now, a leave and not a resignation. But the shape of the wager is clear, and it is the rare AI pitch that gets more convincing the more you learn about the person making it.