From enterprise search to the quote desk

Thirty seconds between pages, two days before quarter-end

Two days before quarter-end, a salesperson is trying to close a deal. The quoting software takes 30 seconds to move between pages, while getting colleagues to process a quote within a day or two requires repeated requests. This opening recollection captures a practical bottleneck: creating and approving an offer can obstruct selling it.

Roadrunner aims to simplify that work by rebuilding the data model behind quotes and using large language models (LLMs) to reason over commercial information. The companion article, published December 12, 2025, identifies it as a Kleiner Perkins incubation led by Joubin Mirzadegan. The proposed combination is a more flexible foundation for pricing rules, enterprise workflows, and eventually recommendations about how to structure deals.

A podcast built to solve a hiring problem

Before getting to enterprise software, the conversation turns to interviewing. Asked about an earlier interview with Mark Zuckerberg, the interviewer describes its purpose as raising the profile of the Chan Zuckerberg Initiative and Priscilla Chan’s research ambitions among engineers. Extensive preparation by its staff reduced the nerves. The interviewer also offers a strongly optimistic judgment: achieving even half of Chan’s ambitions could give her greater impact than Zuckerberg, with Facebook funding the research. That is a conditional assessment of the ambitions.

The podcast creator recalls shifting from chief revenue officers, or CROs, to founders and CEOs around episode 70 or 80. Founders seemed freer to speak without worrying about a boss’s reaction. Their public profiles also attracted listeners, the other participant adds, while pointing to the candid Emily Choi conversation about politics as an example of what a non-CEO interview could achieve.

The show began with a hiring need. The creator recounts a career in startup sales, followed by an acquisition by Palo Alto Networks and an assignment to build its public cloud business in the central United States. Kleiner Perkins then approached them about helping founders sell what they had built. Working with Arvind at Glean exposed the limit of advice: someone had to own sales execution, and finding that person required a network the creator did not yet have.

Interviewing CROs offered an exchange: help operators tell their stories while building relationships that could support future hiring. Kleiner Perkins (KP) was skeptical, preferring its portfolio companies to speak for the firm. A pilot with the creator’s former boss made the idea tangible. After ten episodes showed promise, the creator committed to 100 episodes before judging the project, giving it room to develop without reacting to every criticism. Over roughly six years, production moved from every other week to weekly. The eventual stopping rule remained personal: stop when conversations with guests no longer feel worth looking forward to.

Make room for an unrehearsed conversation

Video let listeners watch while cooking and switch to background audio on a run. For a technical show, it also enabled code, diagrams, and demonstrations, though editing those visuals added cost. Cameras introduced another tradeoff: guests could become self-conscious and begin performing a polished public persona.

The creator’s response was to shape the room around comfort: an attentive greeting, water, cool air, upward-facing lights, and conversation that begins without a formal readiness cue. Preparation happens beforehand through extensive reading. During the interview, the creator probes a mental model of the guest, normally withholding questions in advance to avoid rehearsed answers.

The participants disagree about how absolute that rule should be. One initially says advance screening would mean declining the interview, then starts considering a Zuckerberg exception. The other argues that screening can be the price of access while still allowing unscripted follow-ups. Preparation and spontaneity can coexist; the tradeoff is how much control to concede before sitting down.

Glean: permissions first, organizational trust next

Enterprise search means finding information across an organization’s data. Glean faced both a technical challenge and buyers weary of decades of promises, including from Google. Crawling the data was only part of the problem. Search had to preserve authentication and layered permissions so that useful retrieval never exposed restricted material, such as another employee’s compensation.

Working technology still left deployment and purchasing barriers. Customers had to decide between running it on their premises or in the cloud and assess the security implications of indexing sensitive information. In the guest’s recollection, Glean’s later prominence in AI was far from obvious while the team worked through these obstacles.

Rubrik became the core design partner—an early customer working closely with the team to develop the product. Arvind had encountered the search problem there as a co-founder before leaving to start Glean. The relationship provided access to people and systems, close collaboration, and eventually a production reference for skeptical prospects. The bet was that Rubrik’s requirements would generalize. Access made the work easier, but implementation remained difficult.

Enterprise selling also requires helping customers navigate their own organization. Stakeholders must understand the product, agree on its implications, and move it through approvals. The guest compares this to passing legislation; the interviewer emphasizes mapping champions and resistance.

AI adds two simultaneous decisions: how to use LLMs within the organization and where a particular vendor’s product fits. Forward deployed engineering addresses both by placing vendor engineers in customer environments to advise on useful applications and co-develop a solution. The guest views that substantial technical support as necessary during early enterprise adoption.

A brief career discussion puts people ahead of the technology. One participant ranks prospective employers by people, product, then money: products can change more readily than people, and compensation cannot make up for disliking both. The other doubles the emphasis on people and places market ahead of product—a distinction the first accepts.

Windsurf built distribution alongside the product

KP’s support operation grew from customer relationships and sales advice into dedicated help with sales, access to major customers, and marketing. The aim was to help technical founders who understood product and engineering but had little experience closing deals or generating demand. Windsurf illustrates the importance of building those capabilities deliberately.

The guest attributes Windsurf’s growth to three reinforcing conditions: - Google-class product quality and Salesforce-class distribution: founders committed to engineering and sales from the outset. - Strong demand for AI coding assistance: expensive engineering work and code’s structured nature made assistance an attractive use case. - Fast hiring: the company moved quickly to bring in strong people. The commercial lesson is that individual adoption does not automatically produce an enterprise purchase or expansion.

One concrete investment stood out to the interviewer: an entire floor of Windsurf’s Redwood City office devoted to video production. By recording time, its go-to-market organization—the team responsible for reaching and selling to customers—was at Cognition. The interviewer saw that sales capacity as complementary to Cognition’s product and engineering strengths. The guest credited Liam with substantial work building the organization and Liam, Lauren, and Suzanne with much of KP’s contribution, describing their own role as encouragement and customer introductions.

Hire for the selling conditions you have

Stage fit matters more than employer prestige. Success at Snowflake or Databricks does not by itself establish suitability for an early AI company. Someone who won customers for a third-place product may have more relevant experience than a seller accustomed to brand recognition, inbound leads, and an established playbook. The distinction is situational: mature-company experience can be valuable without transferring automatically to a seed-stage startup.

The proposed interview examines three dimensions: - Resilience: what sustains the candidate through repeated rejection and startup setbacks? - Technical curiosity: have they learned enough to explain a product and its place in a customer’s systems? - Stage experience: have they found ways to win business before the brand, leads, and playbook existed? The technical bar is practical: ask engineers useful questions, understand their answers, and explain the product without relying on a sales engineer for every conversation.

The interviewer cautions that a candidate has not yet received product training or time with founders. The response is to look for a demonstrated willingness to develop that depth elsewhere. A portfolio anecdote supports the emphasis on potential: across five executive roles at eight leading KP companies, the guest reports that 38 of 40 executives were reporting directly to a CEO for the first time. The lesson drawn is to value company context, trust, and learning speed; this is a portfolio observation, not a general hiring success rate.

Early selling requires creative problem-solving. At Windsurf’s more developed stage, the work had become highly systematized. The guest even doubts their own fit for following that mature playbook criterion by criterion. Building a sales process and operating it demand different strengths.

Roadrunner: pricing complexity reaches the data model

Roadrunner returns the conversation to quoting software, called CPQ throughout the discussion: systems used to assemble commercial offers and obtain approval. A simple subscription ties a price to a number of seats. Enterprise catalogs add dozens of products, volume-dependent prices, discounts, renewals, early renewals, and expansions across product lines. These conditions interact to determine which offers a seller can make.

Consumption pricing adds charges based on usage—for example, a minimum spending commitment with additional usage billed above it. Token consumption supplies the AI example. The guest predicts more such arrangements and argues that legacy quoting data models struggle to represent their combinations. Roadrunner’s thesis therefore starts with the commercial rules underneath the interface.

The opportunity surfaced repeatedly in a network of chief information officers (CIOs) that the guest started with Lauren at KP. At one dinner, roughly five CIOs named CPQ as their biggest company problem. Six months later, a different group gave the same answer: revenue leaders and salespeople blamed IT for software that could not support their work. An investment search followed but produced no compelling candidate. GPT-3.5 then prompted the idea that LLMs could reason over structured and unstructured commercial information, with enterprise features and workflows built around them.

The guest compares the required rebuild to the shift from on-premises applications to cloud-native architecture: new technology can change the assumptions on which an application rests. The competitive thesis also depended on a perceived Salesforce product transition, judged at recording time to create roughly a two-year opening. The guest preferred trying to outpace Salesforce to competing with OpenAI. Two barriers made the opportunity difficult: sufficient technology and enterprise distribution, especially among larger customers with the most complicated pricing.

Test the model against difficult customer rules

The podcast and CIO network now supplied relationships on both sides of the problem: revenue leaders felt the pain, while information officers supplied the software. Those connections could recruit design partners and help the startup through legal and procurement processes. After discussing the personal commitment with their partner at home and partners at KP, the guest decided to build the company with technical co-founders.

At recording time, Roadrunner reportedly had nine people, including co-founders AJ and Eugene. They had met at Caltech, worked at Robinhood and Meta respectively, and worked together on Mars rover software at NASA. Their earlier project, Athena, used an LLM to evaluate college applications. Seeking a market that better matched their ambitions, they independently found CPQ by asking founder friends about their biggest problems. The founding team paired their technical ability with enterprise sales experience and customer access.

Roadrunner was co-developing with four design partners through a shared Slack channel and weekly stand-ups. Following Glean’s Rubrik precedent, the team bet that their requirements would generalize. It deliberately sought difficult combinations of hardware, software subscriptions, consumption pricing, and many product identifiers, or SKUs. The aim was to expose problematic combinations before launch; unlimited flexibility remained an aspiration.

A concrete requirement is conditional quoting permission. An account executive—a salesperson assembling the offer—in the UK may be allowed to quote only certain SKUs, with specified discounts and a maximum discount ceiling. Seller location, product eligibility, and pricing authority must therefore work together. Much of development went into testing the data model against real rules and their combinations, including requirements associated with doing business out of Canada.

The guest’s view of expansion follows the same discipline: solve a difficult problem for a specific group of customers, then earn the opportunity to solve adjacent problems. That mission is worthwhile without inflating its social significance. A growing company can also give engineers ownership of entire product initiatives, opportunities to manage, and financial rewards. Expansion follows execution.

From capturing quotes to recommending deals

The longer-term LLM vision depends on routing all of a customer’s quotes through Roadrunner. With that history available, the system could recommend how to structure the next offer. The guest imagines comparing a proposed Costco deal with a similar past Nordstrom deal, then suggesting adjustments, bundling, and presentation. The counterparties are hypothetical; the mechanism is using accumulated deal history to advise on a new transaction.

Today, that knowledge often sits with people. A new representative asks finance, experienced account executives, and deal desk—the staff helping assemble and approve offers—how to put a deal together. The interviewer describes Cognition’s coordination as a busy Slack channel full of requests. Roadrunner’s proposed value is to reduce this administrative back-and-forth so sales and deal-desk staff can spend more time on strategic work. Proactive recommendations were a future ambition at recording time.

The immediate constraint was engineering capacity. The guest recounts wanting to admit ten interested customers while the co-founders objected that the team lacked bandwidth. Customer needs were driving the roadmap. The overall strategy felt clear, but tactics would change, and recruiting people willing to endure early company building remained essential.

Reduce friction, preserve enthusiasm

The closing discussion turns to sustaining the person doing the work. The guest’s defaults are exercise immediately after waking and a salad for lunch. Monday means cycling up Hawk Hill; other days mix running, weights, and sport, recently basketball. The commitment stays fixed while most exercise choices remain flexible. These habits reduce repeated decisions about whether and when to act, though founder demands still make time difficult to protect.

Delegation freed more capacity. After more than 200 podcast episodes, the editor challenged the creator’s habit of reviewing every release and sending extensive notes. Trusting the editing team removed a production bottleneck. Listening was selective too: follow interesting guests across shows, then sample more episodes if the interviewer is compelling. Recommendations included selected episodes of Invest Like the Best and interviews by Joe Rogan, Tim Ferriss, and Shane Parrish.

Sitting on the answering side of the table revealed how much control an interviewer has through questions and a plan. That reflection leads to the guest’s preferred definition of sales: transferring enthusiasm between people. It applies to recruiting, speaking, and interviewing. Communication training should help listeners feel that enthusiasm while preserving its natural character.

The show’s name came from Angela Duckworth’s Grit, and its creator recalls traveling to Pennsylvania to interview her. Paraphrasing her definition as passion plus perseverance sustained over time, the creator puts the emphasis on passion: choosing work and relationships that matter deeply makes persistence feel lighter and enthusiasm easier to convey.

The final question turns back toward technical learning. Asked for a dream guest, the interviewer names Karpathy, valuing his teaching and clarity about what LLMs can do. The ideal is understanding through reconstruction: teach yourself a system by building a version of it, then explain it simply. The guest closes by reflecting on the unfamiliar experience of having the interview table turned.