The useful machine-learning system is rarely the one with the cleanest demo. It is the one that survives a Tuesday afternoon in a crowded airport, notices that one concourse has absorbed an unexpected wave of passengers, and gives a cleaning team a sensible next move. Junaith Ahemed Shahabdeen has spent much of his career in that untidy handoff between a physical signal and a human decision. As CEO and co-founder of Zan Compute, he works on buildings from the least theatrical end: traffic, supplies, service tickets, inspections and the quiet arithmetic of what needs attention now.
That choice of problem says a good deal about Shahabdeen's taste. Smart-building brochures tend to favor gleaming towers and blue streams of data. Facility work happens among paper dispensers, waste bins, aging software and shifting crews. Every elegant prediction eventually meets a cart, a contract and a clock. Shahabdeen built Zan Compute around that meeting point.
His path there was neither a sudden pivot into fashionable AI nor a founder discovering sensors for the first time. Public career records place him at DigitalSun as a software programmer in the early 2000s, then at Lusora as a software architect. In late 2005 he joined Intel, where he worked for nearly nine years as a machine-learning researcher. The titles changed; the recurring question did not. How can machines interpret imperfect signals from the world around them and turn those signals into something useful?
A researcher learns to respect the signal
At Intel, Shahabdeen appeared on research about mobile sensing, activity inference, edge computation and context-aware systems. He co-authored work on iMote2, a wireless sensor platform designed to push serious computation closer to where data is collected. He also worked on Bayesian approaches to recognizing activity and on methods for getting mobile devices to make smarter use of their connections. Those projects belong to an earlier era of connected computing, before every product announcement acquired an AI suffix. Their constraints now feel contemporary: limited power, noisy observations, missing data and the need to extract meaning without pretending the world is tidy.
In 2014, work he co-authored on recommending group activities from location-based social-network data received a Best Paper Award at an ACM SIGSPATIAL workshop. The subject was different from building maintenance, but the habit was familiar. Combine several partial signals. Model what is happening. Recommend a useful next action. Shahabdeen was accumulating a technical grammar that would later become a company.
He also accumulated patents. His name appears on Intel inventions involving mobile connections and device awareness. At Zan Compute, he is listed as the inventor on a sensor-equipped smart garbage bin patent. The latter is almost comic in its plainness: a bin is usually noticed only when it has become a problem. Give it a way to report its state and the schedule no longer has to rely on habit. A neglected object becomes an operational signal.
“Interoperability is the name of the game!”Junaith Shahabdeen, on smart-building systems
The room nobody put on the tour
Zan Compute began in 2014 as an intelligent-building venture in Santa Clara. The company's account of its early discovery process includes a wonderfully unpolished scene. Shahabdeen described cloud automation for restroom service to a facilities manager and saw the manager's interest sharpen. The reaction helped point the young company toward cleaning operations. There was no cinematic breakthrough, only a customer recognizing that a routine headache might finally become measurable.
Commercial cleaning was full of information gaps. A fixed route could send a worker to a quiet room while a busy one ran short of supplies. Complaints arrived after the failure. Supervisors knew patterns through experience, but that knowledge was hard to share across shifts or properties. Zan's early platform joined small sensors with cloud analytics and a mobile application. Foot traffic, supply levels, waste capacity, water events and occupant feedback could become prompts for action rather than entries in a postmortem.
The company's Smart Washroom AI Platform promised a move from clock-based cleaning to need-based service. That sounds obvious until one considers the stack required to make it boringly dependable: retrofit hardware, radio connectivity, device management, analytics, integrations, alerts, worker interfaces and a record that the work happened. The machine-learning model is one actor in a crowded cast. Shahabdeen's operator instinct is to keep the rest of the cast onstage.
The business lesson hiding behind the soap
Founders can borrow the structure without buying a single sensor. Start with a costly uncertainty, not a fashionable capability. Make the uncertainty visible. Connect the observation to a decision. Make sure the decision reaches the person who can act. Then capture the outcome so the next recommendation has a memory. Many software products stop at the dashboard. Zan Compute's logic continues to the dispatch and comes back with verification.
The market also imposes discipline. Property operations are sensitive to return on investment. A model that consumes expensive computing but saves no labor, material or downtime is merely an ornate invoice. Shahabdeen has written publicly about what he sees as fragile economics in cutting-edge AI, where one company's subsidized price becomes another company's infrastructure bill. His conclusion is blunt: in real estate and smart buildings, AI features have to produce provable impact and the architecture has to deliver it efficiently.
Conceptual demand pattern, not company performance data. Real spaces rise and fall unevenly; useful operations respond to the signal instead of pretending every hour is equal.
That economic caution sits beside an engineer's suspicion of shortcuts. In a recent note on AI-generated software, Shahabdeen warned that code can compile and impress on day one while quietly bypassing architecture, invariants and failure modes. His memorable line was that speed without structure becomes a production time bomb. It is a revealing anxiety for a founder selling intelligence into physical operations. A broken novelty is embarrassing. A brittle system coordinating work across an airport or campus is expensive.
“Speed without structure isn't velocity - it's a time bomb ready to go off in a production setup.”Junaith Shahabdeen, on AI-generated software
Language models enter a building
Zan Compute's newer work adds language models and specialized agents to the foundation it spent years assembling. The company introduced ZANI as a multimodal platform for cleaning and real-estate operations. Sensor readings can sit beside schedules, work orders, inspections, audits, workforce information and occupant feedback. Agents can help plan tasks, answer questions about history, assemble reports or guide staff through operational information.
The ordering matters. A language model is not being asked to imagine whether a dispenser is empty. A sensor handles the physical claim. The language layer can explain the condition, place it among competing tasks, translate it for a worker or turn the history into a report. Measurement comes first; language becomes the interface and coordinator. It is a more modest role than omniscience and a more useful one.
Airports make the ambition legible. Passenger flows change with gates and flight schedules. A facility system can infer where demand is building, predict service needs, generate tasks for people or machines, and check whether the loop closed. Shahabdeen has described this orchestration as the next step in airport experience. The aspiration extends beyond cleaner rooms toward buildings that can understand context, prioritize and help manage themselves.
Yet his public comments keep adding guardrails to the vision. Systems must interoperate. Data belonging to a building-service contractor should remain in that contractor's own sandbox. Software should be modular enough to evolve. Unit economics should survive actual use. The building may grow more autonomous, but accountability does not disappear into the cloud.
A career built in the handoff
Shahabdeen's story is less about a single invention than a durable point of view. At DigitalSun and Lusora, he wrote and designed software. At Intel, he researched how computing systems could interpret the physical world. At Zan Compute, he brought the signal all the way to an operating decision. The distance between a sensor event and a completed task is where much applied AI either earns its keep or quietly becomes another dashboard nobody checks.
There is also a sly elegance in choosing cleaning as the proving ground. Everyone understands an empty soap dispenser. Nobody needs a futurist to explain an overflowing bin. The success condition is visible, the waste is countable and the worker remains essential. Such plain outcomes leave less room for technological perfume.
After more than a decade running Zan Compute, Shahabdeen is now placing agents on top of a patient accumulation of sensors, workflows and customer constraints. The language of the industry has changed around him: IoT became smart buildings; machine learning became AI; AI acquired agents. His central demand has stayed stubbornly practical. The building must notice something real, decide something useful and help somebody do the work.
Watch and follow
Shahabdeen shares his current thinking through LinkedIn, while Zan Compute publishes product and field updates through its website and company channels. His conversation on smart cleaning and the smart-building blind spot is also available as an ISSA podcast episode on YouTube.