OctaPulse Is Teaching Fish Farms to See—and Building Toward Farms That Can Act
A fish farm is an unusually hostile place to build a vision system: water distorts, animals move, lighting changes and mistakes scale across thousands of living creatures. OctaPulse is starting there on purpose.
Co-founder
Co-founder
The least obvious AI interface is sometimes the most useful
The current AI boom has trained people to think of artificial intelligence as a text box. OctaPulse is building for an environment where there is no reason to type. Its interface to the world is a camera looking at fish, an edge computer interpreting what it sees and, increasingly, machinery capable of turning perception into action.
The company’s first problem is quality inspection in aquaculture. Hatcheries and farms need to identify deformities, assess broodstock and make decisions about which fish should proceed through production. Manual inspection is slow, inconsistent and physically demanding. Automating that step creates immediate value while also generating the visual dataset needed for more ambitious farm automation.
Paul Grech started with the business problem
Paul Grech came to the company after years at Bloomberg and graduate work at Carnegie Mellon. His background is not that of a lifelong fish farmer; it is closer to product, analytics and business systems. Carnegie Mellon’s founder profile describes a personal connection to the ocean through roots in Malta and Puerto Rico and an interest in building technology around food production and marine systems.
That business orientation is useful because robotics companies often die between technical demonstration and real workflow. A model can be accurate in a demo and still fail if it slows operators, requires delicate calibration or costs more than the labor and losses it replaces. Grech’s side of the problem is turning technical capability into something a farm can actually buy and operate.
Rohan Singh brings the machine side
Rohan Singh’s background runs through mechanical engineering, AI and robotics. He studied mechanical engineering at Texas A&M and later AI engineering at Carnegie Mellon, with experience across companies including Tesla, NVIDIA, ASML and Toyota. He grew up in Goa, India, another place where the relationship between people and the water is difficult to ignore.
For OctaPulse, that combination matters because aquaculture automation is not a pure computer-vision problem. Cameras, enclosures, lighting, motion, timing, controls and eventually robotic mechanisms all have to survive a wet industrial environment. Software has to understand the physical machine, not merely classify an image.
Why inspecting a fish is technically hard
Computer vision works best when the world cooperates: fixed lighting, repeatable backgrounds, stable cameras and objects that hold still. Fish farms offer the opposite. Bodies bend and rotate. Water introduces reflections and refraction. Species have different shapes. Juveniles can be small. The same deformation can look different from a new angle. Dirt, bubbles and equipment create visual noise.
That means a high-performing system needs more than a generic object detector. The data collection process becomes a major part of the technology: which angles are captured, how fish are presented, how labels are defined, how uncertainty is handled and how the system behaves when the visual conditions drift from the training set.
The first wedge: faster quality assurance
Y Combinator says OctaPulse’s current system can reduce an inspection that takes roughly five minutes manually to under thirty seconds, with company-reported accuracy above 90 percent. YC also reports a six-figure paid pilot with a major U.S. trout producer and additional farm deployments planned in 2026. Those numbers are promising, but they are company-reported and still early.
The strategic value of the wedge is clearer than the headline metric. Quality inspection is a discrete task with measurable labor, time and production consequences. It gives OctaPulse a place to earn trust inside the farm while collecting real production data. That is a stronger route into robotics than asking a farm to buy a futuristic “autonomous aquaculture” system on day one.
Aquaculture is already industrial scale
The market context is substantial. The United Nations Food and Agriculture Organization reported that aquaculture surpassed capture fisheries in aquatic-animal production in 2022. Global farmed aquatic-animal output is now measured in more than one hundred million tonnes, making aquaculture a central part of the world’s food system rather than a niche alternative to fishing.
As production grows, farms face the same pressures that push automation into other industries: labor constraints, feed costs, biological loss, consistency, traceability and the need to make decisions earlier. The difference is that the production units are living animals whose welfare and health change the economics.
The dataset could become the real moat
OctaPulse says it is building a multi-species dataset that can serve as the “brain” for future autonomous aquafarms. That is the most strategically interesting part of the company. Hardware can be copied. Cameras become cheaper. Vision-model architectures diffuse. A labeled dataset that captures how different species look across ages, farms, diseases, lighting conditions and operational environments can be much harder to recreate.
But a dataset is only defensible if it keeps compounding. OctaPulse has to design deployments so each installation improves future models without creating a maintenance nightmare of incompatible cameras and labels. The technical organization will need rigorous versioning, model monitoring and farm-specific calibration while still moving toward a general system.
From seeing to acting
The long-term roadmap moves from perception into control: automated grading, feeding, health monitoring, breeding decisions and eventually robotics that can execute more farm tasks. That transition is where the difficulty rises sharply. A wrong classification can be reviewed; a robotic action can affect an animal or production process immediately.
Autonomy therefore has to be earned in layers. The sensible path is decision support first, bounded automation second and broader closed-loop control only after the system has enough evidence to know when it should not act. In biological systems, restraint is a capability.
The risks are physical, not rhetorical
OctaPulse will have to prove cross-species performance, ruggedize equipment, control installation costs and show that model accuracy remains useful outside carefully prepared demonstrations. It will also need to integrate into farm workflows where downtime is expensive and technical support may be far away. Fish welfare adds another dimension: automation that improves throughput while increasing handling stress would be a bad trade.
These are exactly the kinds of constraints that make the company interesting. The technology cannot hide behind a dashboard. It has to function in water, around animals, under industrial pressure and in the hands of people who already know how to run farms.
Why OctaPulse is worth watching now
Carnegie Mellon highlighted OctaPulse in its startup ecosystem before the company entered Y Combinator’s Winter 2026 batch, and the company has already moved into paid production testing. It remains early enough that the central thesis is unproven—but far enough along that there is something concrete to evaluate.
The larger idea is powerful: computer vision becomes a sensory layer for an industry, the data from that layer compounds, and robotics gradually turns perception into better physical decisions. If OctaPulse can build that stack without losing sight of farm economics and animal biology, it will be doing more than applying AI to fish. It will be helping define what autonomous agriculture looks like underwater.