Building an enterprise AI company on one "Base of shifting sand" According to leadership, the main challenge for founders today is Paluna Ai.
Today, the Palo Alto-based startup—backed by ex-Google and Meta Engineering veterans—is making a decisive vertical push in the restaurant and hospitality space with today’s launch of Palona Vision and Palona Workflow.
The new offering transforms the company’s multimodal agent suite into a real-time operating system for restaurant operations.
The news marks a strategic pivot from the company’s launch in early 2025, when it was first revealed. 10 million in seed funding To build emotionally intelligent sales agents for the broader direct-to-consumer businesses.
Now, by limiting its focus to a "Multimodal ancestry" The vision for restaurants, Pulona is providing a blueprint for AI builders on how to move forward. "Thin wrappers" To build deep systems that solve physical-world problems from a higher grain.
“You’re building a company on top of a foundation that’s sand,” said co-founder and CTO Tim Hewes, referring to the instability of today’s LLM ecosystem. “So we’ve built an orchestration layer that allows us to change the model on efficiency, fluidity and cost.”
VentureBeat recently spoke in person with Hose and co-founder and CEO Maria Zhang recently—where else? – A restaurant in NYC about technical challenges and the hard lessons learned from their launch, growth and pivot.
New Offering: Vision and Workflow as ‘Digital GM’
For the end user – the restaurant owner or operator – the latest release of Palona is designed to work as an automation. "Best Operations Manager" Who never sleeps.
Pulona Vision uses in-store security cameras to analyze operational signals—such as queue length, table turnover, PEP interruptions, and cleanliness—without the need for new hardware.
It monitors front-of-house measurements such as queue length, table turns and cleanliness, while simultaneously identifying front-of-house issues such as prep slowdowns or station setup errors.
Pulona Workflow complements this by automating multi-step operational processes. This includes managing catering orders, opening and closing checklists, and completing food preparation. By correlating video signals from the vision with point-of-sale (POS) data and staffing levels, the workflow ensures consistent execution across multiple locations.
“The Paluna vision is like giving digital GM everywhere,” Shaz Khan, founder of Tono Pizzeria + Cheesesteaks, said in a press release provided to VentureBeat. “Before they grow and save me hours every week.”
Going vertical: Lessons in domain expertise
Paluna’s journey began with a star-studded roster. CEO Zhang previously served as VP of Engineering at Google and CTO of Tinder, while co-founder Hewes is the co-inventor of LDAP and former Netscape CTO.
Despite this prescription, the team’s first year was a lesson in the need for focus.
Initially, Paluna served fashion and electronics brands "Wizard" And "Surfer dude" Personalities to handle sales. However, the team quickly realized that the restaurant industry presented a unique, trillion-dollar opportunity that was "Surprisingly recession proof" But "gobsmacked" By operational inefficiency.
"Advice to startup founders: Don’t go multi-industry," Zhang warned.
Vertically, the pulona moved from being a "thin" Chat layer for building a "Multisensory information pipeline" It processes vision, sound and text.
This clarification of focus opened up access to proprietary training data (such as prep books and call transcripts) while avoiding typical data scraping.
1. Building on ‘shifting sand’
To accommodate the reality of enterprise AI deployments in 2025—with new, improved models emerging on an almost weekly basis—Plona developed a patent-pending orchestration layer.
Instead of being "Bundle" With a single provider like OpenAI or Google, Palona’s architecture allows them to change models on a dime based on performance and cost.
They use a mix of proprietary and open-source models, including Gemini for computer vision benchmarks and specific language models for Spanish or Chinese fluency.
For builders, the message is clear: Never let the core cost of your product become a vendor dependency.
2. From vocabulary to ‘global models’
The launch of the Paluna vision represents a shift from understanding words to understanding the physical reality of the kitchen.
While many developers struggle to stitch together disparate APIs, Pulona’s new vision model transforms existing store cameras into operational assistants.
is the identity of the system "Cause and effect" In real-time, identifying if a pizza has been reduced due to "Yellowish beige" Color or alert the manager if a display case is empty.
"In words, physics doesn’t matter," Zhang explained. "But actually, I drop the phone, it always goes down … we really want to know what’s going on in this restaurant world.".
3. The ‘muffin’ solution: custom memory architecture
One of the most important technical hurdles Palona faced was memory management. In a restaurant context, memory is the difference between a disappointing interaction and one "magical" One where the agent misses dinner "As usual" Order
The team initially used an unspecified open-source tool, but found that it produced errors 30 percent of the time. "I think advisory developers always turn off memory (on consumer AI products), because it’s guaranteed to break everything," Zhang warned.
To solve this, Pulona created Muffin, a proprietary memory management system named after the Web. "Cookies". Unlike standard vector-based approaches that struggle with structured data, Muffin is architected to handle four distinct layers:
Structured data: static facts such as shipping addresses or allergy information.
Slow-changing dimensions: Loyalty preferences and favorite objects.
Temporal and seasonal memories: adapting to shifts such as preferring cold drinks in July over hot cocoa in winter.
Regional context: Defaults such as time zone or language preferences.
Lesson for builders: If the best available tool isn’t good enough for your specific vertical, you should be prepared to build your own.
4. Reliability through ‘grace’
In the kitchen, AI isn’t just a typo. This is a lost order or safety hazard. A recent event Stefanina’s Pizzeria in Missouri, where AI fakes deals during the dinner rushhighlighted how quickly brand trust can evaporate when safeguards are absent.
To prevent such chaos, Pulona’s engineers follow its internals Framework of grace:
Safeguards: Strict limits on agent behavior to prevent unauthorized promotions.
Red Teaming: Proactive Efforts "break" Identify motivations for AI and possible deception.
App Second: Lock down APIs and third-party integrations with TLS, tokenization, and attack prevention systems.
Compliance: Grounding every response in verified, tested menu data to ensure accuracy.
Escalation: Routing complex interactions to a human manager before the guest receives incorrect information.
This reliability is verified by large-scale simulations. "We simulated a million ways to order a pizza," Using one AI to act as a customer and another to take orders, measuring accuracy to eliminate deception, Zhang said.
The bottom line
With the release of Vision and Workflow, Pulona is betting that the future of enterprise AI is not in broad assistants, but in expertise. "Operating system" One who can see, hear and think in a specific domain.
Unlike general-purpose AI agents, Palona’s system isn’t designed to just answer questions, but to execute restaurant workflows—it’s capable of remembering customers, listening to their orders. "usual," And monitoring restaurant operations to ensure they serve food to that customer in accordance with their internal processes and guidelines, flagging whenever something is wrong or significantly wrong. about go wrong
For Zhang, the goal is to get human operators to focus on their craft: "If you’ve got that delicious meal nailed down … we’ll tell you what to do."