Build an enterprise AI company "moving sand foundation" According to Palona AI leaders, this is a core challenge for today’s founders.
Now led by a former Google and Meta Engineering veteran, the Palo Alto-based startup is making a decisive vertical move into the restaurant and hospitality space with today’s launch of Palona Vision and Palona Workflow.
The new product transforms the company’s multimodal agent suite into a real-time operating system for restaurant operations that spans cameras, calls, conversations, and coordinated task execution.
The news marks a strategic shift from the company’s debut in early 2025, when it first emerged with $10 million in seed funding to build emotionally intelligent distributors for a wide range of direct-to-consumer companies.
Now, by narrowing our focus, "multimodal native" Parona offers AI builders an approach that goes beyond that for restaurants "thin wrapper" Build deep systems to solve high-stakes physical world problems.
“We’re building a company on a foundation of sand. Not quicksand, but shifting sand,” said co-founder and CTO Tim Howes, referring to the instability of today’s LLM ecosystem. “So we built an orchestration layer that allows models to be swapped out depending on performance, fluency, and cost.”
VentureBeat recently spoke directly with Howes and co-founder and CEO Maria Zhang. Where else? — at a New York restaurant to discuss the technical challenges and tough lessons learned from launching, growing, and pivoting.
New service: Vision and workflow as a “digital GM”
For end users (restaurant owners and operators), the latest release of Palona is designed to function as an automated system. "best operations manager" That is to never sleep.
Palona Vision uses in-store security cameras to analyze operational signals such as line length, table turnover, prep bottlenecks, and cleanliness without the need for new hardware.
Monitor front-of-house metrics such as queue length, table turnover, and cleanliness, while also identifying back-of-house issues such as late preparation or station setup errors.
Palona Workflow complements this by automating multi-step operational processes. This includes managing catering orders, starting and closing checklists, and fulfilling meal preparation. By correlating video signals from Vision with point of sale (POS) data and staffing levels, workflows ensure consistent execution across multiple locations.
“Parona Vision is like giving digital GM everywhere,” said Shaz Khan, founder of Tono Pizzeria + Cheesesteaks, in a press release provided to VentureBeat. “It saves me time each week by alerting me to problems before they escalate.”
Going Vertical: Lessons from Domain Expertise
Parona’s journey began with a star-studded roster. CEO Zhang previously served as Google’s vice president of engineering and Tinder’s CTO, and co-founder Howes is the co-inventor of LDAP and former Netscape CTO.
Despite this pedigree, the team’s first year was a lesson in the need for focus.
Initially, Parona served fashion and electronics brands, "wizard" and "surfer man" Personality responsible for sales. However, the team quickly realized that the restaurant industry presented a unique multi-trillion dollar opportunity. "surprisingly resistant to recession" but "gobmac" By operational inefficiency.
"Advice for startup founders: Don’t go into multiple industries." Zhang warned.
Through verticalization, Parona has become more than just a company. "thin" Build the chat layer "Multisensory information pipeline" Process vision, audio, and text together.
This clarity of focus enabled us to access our own training data (such as preparation playbooks and call transcripts) while avoiding common data scraping.
1. Build on “shifting sands”
To address the realities of enterprise AI adoption in 2025, with new and improved models emerging almost every week, Parona developed a patent-pending orchestration layer.
rather than being "bundled" When using a single provider like OpenAI or Google, Palona’s architecture allows you to exchange models for 10 cents based on performance and cost.
They use a combination of proprietary and open source models, such as Gemini for computer vision benchmarks and specific language models for Spanish or Chinese fluency.
For builders, the message is clear. Don’t make your product’s core value a dependency on a single vendor.
2. From words to “world models”
The launch of Palona Vision represents a transition from understanding words to understanding the physical reality of the kitchen.
While many developers struggle to piece together separate APIs, Parona’s new vision model turns existing in-store cameras into operational assistants.
system identify "cause and effect" Recognizes in real time if your pizza is undercooked. "light beige" Change the color or alert the administrator if the display case is empty.
"So physics doesn’t matter." Zhang explained. "But actually, I drop my phone and it hangs up all the time… We really want to know what’s going on in this restaurant world".
3. “Muffin” Solution: Custom Memory Architecture
One of the most important technical hurdles Palona faced was memory management. In a restaurant context, memory can make the difference between an irritating interaction and an unpleasant one. "magical" If the agent remembers the diner "usual" order.
The team initially used an unspecified open source tool, but found it had a 30% error rate. "I think advisory developers always turn off memory (in consumer AI products). Because that’s guaranteed to ruin everything." Zhang warned.
To solve this, Palona built Muffin, a proprietary memory management system named after the web. "cookie". Unlike standard vector-based approaches that struggle with structured data, Muffin is designed to handle four different layers.
-
Structured data: Stable facts such as shipping addresses or allergy information.
-
Slowly changing dimensions: Loyalty preferences and favorite items.
-
Temporary seasonal memory: Adapts to change, such as preferring cold drinks in July and hot cocoa in winter.
-
Regional context: Defaults such as time zone and language settings.
Lesson for builders: If the best available tools aren’t good enough for your particular industry, you should build your own.
4.Reliability with “GRACE”
In the kitchen, AI errors are more than just typos. That’s either a wasted order or a safety risk. A recent incident at Stefanina’s Pizzeria in Missouri, where AI created hallucinations of fake transactions during the dinner rush, highlights how quickly trust in a brand can evaporate if safety measures are not taken.
To prevent such confusion, Palona engineers follow the internal GRACE framework.
-
Guardrails: Tight restrictions on agent behavior to prevent unauthorized promotions.
-
Red Teaming: An Aggressive Attempt "break" Use AI to identify potential hallucination triggers.
-
App Sec: Lock down API integration with third-party TLS, tokenization, and attack prevention systems.
-
Compliance: We base all answers on verified and vetted menu data to ensure accuracy.
-
Escalation: Route complex interactions to human managers before guests receive incorrect information.
This reliability is verified through extensive simulations. "We simulated a million ways to order pizza." Zhang said one AI will act as a customer and the other will be used to fulfill orders, measuring its accuracy in eliminating hallucinations.
conclusion
With the launch of Vision and Workflow, Parona is betting that the future of enterprise AI lies in specialized rather than broad assistants. "operating system" You can see, hear, and think within a particular area.
In contrast to general-purpose AI agents, Parona’s system is designed to not only respond to queries but also execute a restaurant’s workflow, remembering customers and listening to their orders. "The usual," And it’s about monitoring the restaurant’s operations to ensure it’s serving customers according to internal processes and guidelines, and flagging any issues or major issues that arise. About I’ll make a mistake.
Zhang’s goal is to free up human operators to focus on their work. "If you’re hooked on delicious food…we’ll show you what to do."
