Late replenishment
Cooking starts when a tray is already low, so the batch lands after the pan is empty.
The opportunity
Empty trays are discovered by guests, not by the kitchen.
Buffet service still runs on walk-past checks. By the time a steward notices a low pan, the next batch is twenty minutes away, and the guest has already seen the gap. SmartBuffet AI turns each tray into a live signal the pass can act on.
Looks full, is not
A well-arranged platter can hide how little food remains. The eye is not a scale.
Safety on paper
Holding temperatures and corrective actions live in clipboards rather than a timestamped record.
No service memory
Nobody can say how long a tray stood empty, or how fast the kitchen answered.
The concept
Measure, see, protect, then decide.
Each technology answers the question it is best at. The decision layer combines them into one instruction for the kitchen.
Smart weighing
A load cell under each pan reports food weight with the container tared out. This is the source of truth for quantity, pace, and refills.
AI vision
One overhead camera covers several platters, identifies each dish, and flags bare or misleading trays. It explains; it does not weigh.
Temperature
Probes on hot and liquid dishes watch the holding band. A full tray never clears a temperature problem.
Kitchen intelligence
Pace, prep time, and status become one instruction: start the batch, refill now, or check the probe.
The hardware
AI vision, weight and temperature in one station unit.
The SmartBuffet monitoring unit combines a high-precision load-cell platform, an overhead AI camera, and an optional temperature probe. Live weight, dish recognition and holding temperature feed the kitchen display in real time.
High-precision scale
Load cells with tare support and about ±5 g accuracy. Food weight only — the container is excluded.
AI vision camera
Detects the dish, estimates fill level and flags empty or low trays. One camera can cover a station zone.
Temperature probe
Optional holding-temperature monitoring keeps hot dishes inside the safe band, separate from weight.
Live demonstration
See it work before any hardware is bought.
The kitchen display runs a full breakfast station in the browser. Every rule shown on this page is working on that screen, with simulated sensors standing in for real ones.
- Load the biryani example and watch the system call for the next batch.
- Serve fried rice three times and see a stock-out predicted at half a tray.
- Refill the soup and watch a temperature warning open on a full pan.
- Read a tray on the pad and see food separated from tare.
- Leave a critical alert unanswered and watch it escalate.
Innovation
What makes this more than a scale with a screen.
Every capability below is working in the proof of concept today.
Predictive replenishment
Compares time-to-empty with kitchen prep time and asks the pass to start cooking while there is still cover, not after the tray is low.
Sensor-fusion trust
When the camera sees a fuller tray than the scale measures, the display says so and trusts the weight. Disagreement is shown, never hidden.
Event intelligence
Every weight change is classified: guest serving, refill, top-up, or a utensil spike. A refill is never mistaken for negative consumption.
Cook messaging and escalation
Alerts go to the cook assigned to that tray. One alert per condition, and an unanswered critical alert escalates to the supervisor.
Tray reading pad
A placed tray settles into food, tare, and gross before a new full batch is confirmed. Unstable readings are rejected.
Station digital twin
A live map of the counter shows each tray's fill, status, and temperature at a glance, from the pass or a tablet.
Service readiness
Before doors open, the station reports how many trays are full, in temperature, and online. Service starts from a verified state.
Empty-tray time
The system counts the minutes guests faced a bare tray. It is the clearest single measure of buffet service quality.
Batch age and chef approval
Each tray shows how long its batch has been out. The AI recommends the next batch; the chef approves or holds it.
Predictive replenishment
Start cooking when the numbers say so.
A percentage tells you where a tray is. Pace tells you where it is going. SmartBuffet AI compares the two with the time the kitchen needs.
Chicken biryani
Replenishment predictedRemaining food2.4 kg
Consumption rate0.8 kgper 10 minutes
Estimated time to empty30 min
Kitchen preparation time25 min
Kitchen action required
Start preparing the next batch now. Starting now leaves about 5 minutes of cover before the tray is empty.
Time to empty = remaining food ÷ consumption rate
2.4 kg ÷ 0.08 kg per minute = 30 minutes.
Start the batch when time to empty ≤ prep time + 5 min cover
30 minutes is inside 25 + 5, so the kitchen is asked to begin. The chef approves the batch; the system never cooks or changes the weight on its own.
The rate uses the servings in the last fifteen minutes. A pilot would tune this window and the cover margin on real service data.
Measurement
Food weight only. Never the pan.
A pineapple platter reads 4,400 g gross on a 1,200 g tray. The system reports 3,200 g of fruit, which is 80% of the 4,000 g full batch.
Remaining % = current food weight ÷ initial food weight × 100
| Remaining | Status | Kitchen response |
|---|---|---|
| Above 60% | Normal | Continue monitoring |
| 41–60% | Monitor | Visible on the card |
| 21–40% | Replenish soon | Schedule the next batch |
| 1–20% | Low stock | Critical: prepare now |
| 0% | Empty | Critical: refill immediately |
| Dish | Sensing | Full batch |
|---|---|---|
| Pineapple slices | Weight + camera | 4.0 kg |
| Watermelon | Weight + camera | 5.0 kg |
| Chicken biryani | Weight + probe | 8.0 kg |
| Fried rice | Weight + probe | 7.0 kg |
| Vegetable noodles | Weight + probe | 6.0 kg |
| Tomato soup | Weight + probe | 5.0 kg |
Thresholds are configurable per dish. These values are starting points for the pilot, agreed with the executive chef.
AI vision
One camera, many trays.
An overhead camera draws a region around each platter, recognises the dish, and compares what it sees with what the scale measures. When the two disagree, the weight wins and the screen says why.
Pineapple · 80%
Watermelonscale 42% · camera 66% Papaya Orange Mixed fruit Watermelon looks fuller than it isWeight says 42% · Fruit steward notified
Dish recognition
Each region is tied to a registered dish, so a swapped platter is noticed.
Presentation check
Bare corners, spills, and uncovered food are flagged for a steward to verify.
Privacy by design
Cameras frame the food, not the guests. No face recognition is used.
Replenishment lifecycle
From a falling number to a closed task.
Detect
The scale crosses a band or the pace shortens.
Predict
Time to empty is compared with prep time.
Notify
The assigned cook receives one alert.
Acknowledge
Response time is recorded by role.
Refill
Weight rises; the event is logged as a refill.
Verify
Status resets; temperature is checked separately.
Food safety alignment
Quantity and safety are separate signals.
A full tray can still be unsafe. SmartBuffet AI keeps temperature, time, and corrective action as their own record, supporting the hotel's HACCP plan rather than replacing it.
Working in the POC
- Holding temperature on hot and liquid dishes.
- Warning when a refilled dish is below the holding line.
- Corrective action recorded with the person and time.
- Batch age shown for every tray.
Reference, not regulation
The demo uses 57°C, the FDA Food Code 2022 hot-holding figure. The pilot adopts the limits in the hotel's approved HACCP plan and applicable Sri Lankan regulations. Sensors support staff judgement; they do not declare food safe.
Architecture
Edge first. Cloud when it adds value.
The station keeps working through a network outage. Sensors report to a local gateway, and the kitchen display never depends on the internet to show an empty tray.
SensingLoad cells with an industrial ADC, food-safe temperature probes, PoE overhead cameras.
Edge gatewayESP32 controllers to a local gateway; camera inference on an NVIDIA Jetson-class device; local buffering.
Real-time platformMQTT messaging, NestJS services, PostgreSQL with TimescaleDB for the event ledger.
ExperienceKitchen display, supervisor mobile alerts, and a management view, built in Next.js.
Proof of concept todayThe full station logic runs in the browser and publishes the same station message a real gateway would send.
HTL01 / BRK-01 / HOT-TRAY01 → { deviceId, foodG, tareG, tempC, status, ts }
Designed for the pilot
Next ideas, ready to build on the same station.
Smart pan identity
An NFC or QR tag on each pan binds the dish, tare, and batch the moment it touches the platform. No manual registration.
Batch rotation alerts
A hotel-set time limit per dish prompts staff to swap in a fresh pan rather than top up an old one.
Service close-out
Remaining weight is recorded per tray at the end of each service, the foundation for waste analytics later.
Allergen-linked labels
Each registered dish carries its allergens, ready for staff tablets and guest-facing displays.
Daily service report
Empty-tray minutes, response times, refills, and temperature excursions summarised for the executive chef.
Multi-station scale-out
The device naming runs hotel to sensor, so twenty trays grow to two hundred without redesign.
Pilot and acceptance
Proven on a real counter, measured against a reference.
Demonstrated in the POC
- Food percentage excludes tare.
- Status bands and alerts follow the configuration.
- Refills are recognised and reset availability.
- Utensil spikes are ignored.
- An offline sensor is reported and its last reading held.
- Stock-out is predicted ahead of the low-stock band.
- Unanswered critical alerts escalate.
Validated in the pilot
- Load cell accuracy against a calibrated reference scale.
- Readings at 20, 40, 60 and 80 percent removal.
- Pan swaps absorbed by tare, not counted as food.
- Camera recognition on the hotel's own dishes.
- Alert delivery time on the hotel network.
- Reduction in empty-tray minutes against a baseline.
Acceptance targets are agreed during the site assessment, after the hotel's own dishes and counters have been seen.
One instrumented tray
Real load cell and probe, verified against a reference scale, including pan changes and utensils.
One live station
A full breakfast service with kitchen staff using acknowledge and refill on the display.
Hotel-wide rollout
Expand to more stations and outlets once accuracy and adoption are proven on the first.
See Station 01 in action.
A working kitchen display with prediction, vision, and safety signals, ready to demonstrate.