Constructing an enterprise AI firm on a “basis of shifting sand” is the central problem for founders at this time, in accordance with the management at Palona AI.
At present, the Palo Alto-based startup—led by former Google and Meta engineering veterans—is making a decisive vertical push into the restaurant and hospitality house with at this time’s launch of Palona Imaginative and prescient and Palona Workflow.
The brand new choices remodel the corporate’s multimodal agent suite right into a real-time working system for restaurant operations — spanning cameras, calls, conversations, and coordinated process execution.
The information marks a strategic pivot from the corporate’s debut in early 2025, when it first emerged with $10 million in seed funding to construct emotionally clever gross sales brokers for broad direct-to-consumer enterprises.
Now, by narrowing its focus to a “multimodal native” method for eating places, Palona is offering a blueprint for AI builders on how one can transfer past “skinny wrappers” to construct deep techniques that remedy high-stakes bodily world issues.
“You’re constructing an organization on high of a basis that’s sand—not quicksand, however shifting sand,” stated co-founder and CTO Tim Howes, referring to the instability of at this time’s LLM ecosystem. “So we constructed an orchestration layer that lets us swap fashions on efficiency, fluency, and price.”
VentureBeat spoke with Howes and co-founder and CEO Maria Zhang in particular person just lately at — the place else? — a restaurant in NYC concerning the technical challenges and onerous classes realized from their launch, development, and pivot.
The New Providing: Imaginative and prescient and Workflow as a ‘Digital GM’
For the top person—the restaurant proprietor or operator—Palona’s newest launch is designed to operate as an automatic “greatest operations supervisor” that by no means sleeps.
Palona Imaginative and prescient makes use of in-store safety cameras to investigate operational indicators — comparable to queue lengths, desk turnover, prep bottlenecks, and cleanliness — with out requiring any new {hardware}.
It screens front-of-house metrics like queue lengths, desk turns, and cleanliness, whereas concurrently figuring out back-of-house points like prep slowdowns or station setup errors.
Palona Workflow enhances this by automating multi-step operational processes. This contains managing catering orders, opening and shutting checklists, and meals prep success. By correlating video indicators from Imaginative and prescient with Level-of-Sale (POS) information and staffing ranges, Workflow ensures constant execution throughout a number of areas.
“Palona Imaginative and prescient is like giving each location a digital GM,” stated Shaz Khan, founding father of Tono Pizzeria + Cheesesteaks, in a press launch supplied to VentureBeat. “It flags points earlier than they escalate and saves me hours each week.”
Going Vertical: Classes in Area Experience
Palona’s journey started with a star-studded roster. CEO Zhang beforehand served as VP of Engineering at Google and CTO of Tinder, whereas Co-founder Howes is the co-inventor of LDAP and a former Netscape CTO.
Regardless of this pedigree, the group’s first yr was a lesson within the necessity of focus.
Initially, Palona served trend and electronics manufacturers, creating “wizard” and “surfer dude” personalities to deal with gross sales. Nevertheless, the group shortly realized that the restaurant {industry} offered a novel, trillion-dollar alternative that was “surprisingly recession-proof” however “gobsmacked” by operational inefficiency.
“Recommendation to startup founders: do not go multi-industry,” Zhang warned.
By verticalizing, Palona moved from being a “skinny” chat layer to constructing a “multi-sensory data pipeline” that processes imaginative and prescient, voice, and textual content in tandem.
That readability of focus opened entry to proprietary coaching information (like prep playbooks and name transcripts) whereas avoiding generic information scraping.
1. Constructing on ‘Shifting Sand’
To accommodate the truth of enterprise AI deployments in 2025 — with new, improved fashions popping out on an almost weekly foundation — Palona developed a patent-pending orchestration layer.
Quite than being “bundled” with a single supplier like OpenAI or Google, Palona’s structure permits them to swap fashions on a dime based mostly on efficiency and price.
They use a mixture of proprietary and open-source fashions, together with Gemini for laptop imaginative and prescient benchmarks and particular language fashions for Spanish or Chinese language fluency.
For builders, the message is obvious: By no means let your product’s core worth be a single-vendor dependency.
2. From Phrases to ‘World Fashions’
The launch of Palona Imaginative and prescient represents a shift from understanding phrases to understanding the bodily actuality of a kitchen.
Whereas many builders battle to sew separate APIs collectively, Palona’s new imaginative and prescient mannequin transforms current in-store cameras into operational assistants.
The system identifies “trigger and impact” in real-time—recognizing if a pizza is undercooked by its “pale beige” coloration or alerting a supervisor if a show case is empty.
“In phrases, physics do not matter,” Zhang defined. “However in actuality, I drop the cellphone, it at all times goes down… we need to actually determine what is going on on on this world of eating places”.
3. The ‘Muffin’ Resolution: Customized Reminiscence Structure
Some of the important technical hurdles Palona confronted was reminiscence administration. In a restaurant context, reminiscence is the distinction between a irritating interplay and a “magical” one the place the agent remembers a diner’s “standard” order.
The group initially utilized an unspecified open-source software, however discovered it produced errors 30% of the time. “I feel advisory builders at all times flip off reminiscence [on consumer AI products], as a result of that may assure to mess every little thing up,” Zhang cautioned.
To resolve this, Palona constructed Muffin, a proprietary reminiscence administration system named as a nod to internet “cookies”. In contrast to customary vector-based approaches that battle with structured information, Muffin is architected to deal with 4 distinct layers:
-
Structured Information: Steady info like supply addresses or allergy data.
-
Sluggish-changing Dimensions: Loyalty preferences and favourite objects.
-
Transient and Seasonal Recollections: Adapting to shifts like preferring chilly drinks in July versus scorching cocoa in winter.
-
Regional Context: Defaults like time zones or language preferences.
The lesson for builders: If the very best accessible software is not ok on your particular vertical, you have to be prepared to construct your personal.
4. Reliability by way of ‘GRACE’
In a kitchen, an AI error is not only a typo; it’s a wasted order or a security threat. A latest incident at Stefanina’s Pizzeria in Missouri, where an AI hallucinated fake deals during a dinner rush, highlights how shortly model belief can evaporate when safeguards are absent.
To stop such chaos, Palona’s engineers comply with its inside GRACE framework:
-
Guardrails: Onerous limits on agent conduct to forestall unapproved promotions.
-
Pink Teaming: Proactive makes an attempt to “break” the AI and establish potential hallucination triggers.
-
App Sec: Lock down APIs and third-party integrations with TLS, tokenization, and assault prevention techniques.
-
Compliance: Grounding each response in verified, vetted menu information to make sure accuracy.
-
Escalation: Routing advanced interactions to a human supervisor earlier than a visitor receives misinformation.
This reliability is verified by way of huge simulation. “We simulated one million methods to order pizza,” Zhang stated, utilizing one AI to behave as a buyer and one other to take the order, measuring accuracy to eradicate hallucinations.
The Backside Line
With the launch of Imaginative and prescient and Workflow, Palona is betting that the way forward for enterprise AI is not in broad assistants, however in specialised “working techniques” that may see, hear, and assume inside a particular area.
In distinction to general-purpose AI brokers, Palona’s system is designed to execute restaurant workflows, not simply reply to queries — it is able to remembering clients, listening to them order their “standard,” and monitoring the restaurant operations to make sure they ship that buyer the meals in accordance with their inside processes and tips, flagging every time one thing goes unsuitable or crucially, is about to go unsuitable.
For Zhang, the purpose is to let human operators deal with their craft: “In case you’ve bought that scrumptious meals nailed… we’ll inform you what to do.”
