The seats this robot takes are the ones people don't want: −4 to +4 °C, wet, full sanitary gowning, no phone on the floor, and turnover to match. Necessary work — and the wrong work for a person to be doing all day.
We deploy operated robotic portioning systems onto real production lines, capture the data that only appears when a real robot meets real food at production speed, and curate it into reliability gains no off-floor model can reproduce.
Merista's first product is an operated portioning robot for chilled prepared-meal lines: perception- and weight-guided deposits, a food-safe arm, inline QA, every portion logged, in the footprint of one worker station — operated by Merista on the customer's line. Repetitive portioning is painful, budgeted, and provable: real operational pain with a real budget attached today. That's the wedge. The commercial story lives on the main page. This page is about where the wedge goes.
The seats this robot takes are the ones people don't want: −4 to +4 °C, wet, full sanitary gowning, no phone on the floor, and turnover to match. Necessary work — and the wrong work for a person to be doing all day.
The robot is designed down, not up: no mobility stack, one task family at a time, a one-station footprint, standard gastronorm pans. Every deleted subsystem is uptime and margin — reliability at production speed comes from what we refused to build.
StageWe are selecting design-partner factories in Europe now.
Real food is deformable, variable, and unforgiving — and no off-floor model has met it at production speed. Robotics is at an inflection point: foundation models show real generalization, but reliable deployment still depends on vertical data, task-specific grounding, and production feedback loops. For high-variance food manipulation, that gap closes slowly — and it is gated on physical access, operational trust, and deployment history, not just on model architecture.
The barriers are physical and institutional as much as algorithmic:
Wet, caustic, cold environments are hostile to most hardware. Surviving daily sanitation is a hardware problem before it is a model problem.
Food-contact and retailer-audit certification is a wall we are climbing — and one that keeps generalists out.
Deformable, variable ingredients resist the lab-to-floor shortcut. Grounding comes from the floor, not the bench.
A better model shrinks the common cases. It does not hand anyone a floor position — or a certificate.
Deploy → the robot meets real edge cases, failures, exceptions — missed grasps, texture shifts, SKU changes, line conditions → we capture the interventions and recovery actions → we curate them into grounded supervision → reliability improves → we win more floors → the dataset and the deployment playbook compound.
Capture is the unfair part. This data only exists where our systems run.
Curation is the conversion. Turning raw failure into supervision is an operational capability, not a software feature.
Portioning — repetitive, hard to staff, hygiene-critical, expensive. The paid wedge.
Not selling a robot — taking over the work, task by task: portioning, then topping, inspection, rework, changeovers. One station becomes a line, one line becomes a floor.
Physical and digital agents — robots, sensors, QA, remote operators, planning, recovery, scheduling — converging toward foundation-model-driven autonomy.
Every meal the developed world eats passes through a factory floor still run largely by hand. The destination is to make those floors autonomous: a production floor where atoms and bits move together.
The self-driving food factory.
That is the ceiling and the reason the wedge is worth building — not a deliverable of the current stage.
Seven years building and operating food robotics at REMY Robotics: 10 robot-run kitchens across 5 cities, 15,000 operating hours, 99.9% order completion, <1% human intervention, <15 min mean recovery — in live, customer-facing service.
That record was achieved at REMY Robotics by the Merista founding team. It validates the team's capability. Merista's own systems performance is yet to be proven — that is exactly the work this stage is for.
Including if you're an investor who wants the full story — one address.
hello@merista.ai