As artificial intelligence continues to reshape industries that depend on real-world precision, agriculture has emerged as one of the most complex and high-stakes environments for deployment. Unlike controlled lab settings, farms present constantly shifting variables, weather, terrain, connectivity, and biological unpredictability, making them a proving ground for technologies that must operate reliably at scale.
At the center of this transformation are engineers who can bridge the gap between theoretical innovation and systems that perform in the field every day. Harikrishnan Unnikrishnan, a Senior Perception Engineer, is one of those builders. With a Ph.D. in Electrical Engineering from the University of Kentucky and more than 15 years of experience across edge AI, computer vision, and robotics, Unnikrishnan has consistently focused on translating advanced research into scalable, production-grade systems. His career spans leading engineering efforts that deployed vision-guided robotics across thousands of units, architecting spatial analytics platforms running across more than 10,000 edge devices, and now developing multimodal perception systems designed to bring precision and efficiency to modern agriculture.
In this interview with AI Journal, Unnikrishnan discusses how he moved from academic research into real-world deployment, the engineering realities of scaling AI systems beyond the lab, and what it takes to build resilient, intelligent infrastructure in environments where failure is not an option
To start, how did you first become interested in Edge AI, computer vision, and robotics, and what led you from your Ph.D. into building real-world systems at scale?
As an engineering student, I found the mathematics of signal processing fascinating but abstract. That changed during my graduate studies at the University of Kentucky, where I focused on computer vision and audio signal processing. Applying theoretical derivations to audio and image processing allowed me to finally see and hear the math in action. During my Ph.D. research, I was responsible for the computer vision components of a structured light endoscope designed for high-speed vocal fold recording. I had to bridge the gap between theory and the physical constraints of medical hardware. What led me to building at scale was the thrill of seeing AI interact with the physical world. Whether it was scaling vision-guided robotics to thousands of units at KeyMe or architecting systems for 10,000+ edge devices at Butlr, I’ve stayed driven by the challenge of taking AI from the controlled environment into the hands of customers.
You have successfully transitioned from academic research into deploying production-grade perception systems. What were the biggest gaps you encountered between lab performance and real-world infrastructure, and how did you overcome them?
The primary gap between the laboratory and a production environment is that in research you focus on algorithmic novelty such as the Kalman state-based modeling I utilized for vocal fold kinematics. In a production environment, the engineering innovation lies in the failsafes around the algorithm. Critical features like battery life, serviceability, and data bandwidth consumption are what differentiate a prototype from a scalable product. To overcome these gaps, I focused on building infrastructure that allows AI to manage its own physical constraints. At KeyMe, I architected self-calibrating hardware-in-the-loop systems using AI. This effectively eliminated the need for frequent manual service visits. It reduced error rates by 40%. For a fleet of thousands of units deployed nationwide, this is the key to making the vision-guided system a product that serves thousands of customers daily.
You are designing multimodal perception systems for precision agriculture. How do you ensure these systems remain reliable and accurate across varying environmental conditions and hardware constraints?
Edge AI for precision agriculture is a unique challenge because the hardware is literally ‘in the wild’. They are often in remote environments with minimal internet connectivity and difficult serviceability. In these conditions, reliability is the foundation of the product. I ensure accuracy and uptime through redundancy and modularity. First, by employing multiple modes such as the cameras and GPS integration, we ensure there is more than one way to verify critical spatial data. Second, I’ve moved towards a modular, event-based architecture. This design ensures that the failure of a single non-critical component is not catastrophic to the entire system. Finally, to manage these systems at scale, observability is essential. Having robust log trails and remote monitoring tools—like Grafana and Docker-based infrastructure is the only way to proactively trace and resolve issues before they impact the customer.
During your time at KeyMe Inc., you led deployments of vision-guided robotics across thousands of units, what were the key engineering and operational challenges in scaling to that level, and how did you address issues like latency, uptime, and system consistency?
Scaling vision-guided robotics truly shifts the engineering challenge from algorithmic novelty to systemic reliability. While a 1% error rate might be manageable with 100 units, at a scale involving thousands of units, it represents a catastrophic operational burden. We addressed the critical issue of latency by pushing our computer vision and AI processing directly to the edge.
Maintaining uptime was another massive hurdle. Performing routine manual servicing on thousands of units deployed nationwide was impossible. To solve this, I architected multi-camera vision systems and self-reporting AI that enabled units to proactively report servicing needs. For scaling, the way the technology stack is organized is frequently overlooked. I ensured our architecture remained modularized so that a cross-functional team could innovate swiftly without stepping on each other’s toes.
At Butlr Technologies, you worked on spatial analytics systems running continuously across more than 10,000 edge devices. What strategies did you use to maintain performance, manage data streams, and ensure reliability in always-on environments?
Achieving reliability across 10,000+ edge devices was, of course, a massive team effort. My role as a leader was to provide the architectural North Star. We viewed our occupancy sensors as a nervous system: distributed, intelligent, and intentionally lightweight. One of the most critical strategic decisions I made was to remove as much complexity from the edge as possible. There is a common misconception that Edge AI involves running massive, complex models on the device. Instead, I worked with the team to architect a system in which each unit performs a simplified, highly optimized task. My focus was on the high-throughput data streams to our central analytical systems. We performed essential signal processing and intelligent data conditioning at the edge to improve the Signal-to-Noise Ratio (SNR), while the heavy computational lifting was handled centrally.
Your open-source project, OpenGlottal, combines YOLOv8 and U-Net for clinical waveform extraction. What motivated you to develop this tool, and how does it reflect your broader approach to translating research into accessible, real-world applications?
My commitment to this field began during my PhD at the University of Kentucky, where I focused on analyzing vocal fold kinematics. To this day, I stay deeply engaged with the scientific community by serving as a reviewer for the Journal of Speech, Language, and Hearing Research (JSLHR).
In the last few years, we have seen computer vision AI surpass previous accuracy and efficiency metrics by an order of magnitude. However, there is often a significant lag between these breakthroughs and their adoption in a clinical setting. My motivation for developing OpenGlottal was to leverage my decade of industry experience in scaling AI to deliver a state-of-the-art tool that advances clinical adoption.
OpenGlottal translates complex research into an accessible, real-world application by leveraging recent advancements such as YOLOv8 and U-Net to deliver near-real-time, high-precision segmentation and kinematic tracking. By open-sourcing, I hope to provide clinicians with the tools they need to improve patient outcomes through data-driven insights.
Across your leadership roles, you have built and scaled engineering teams from early stage to national deployment. How do you structure teams and processes to successfully move Edge AI systems from prototype to large-scale infrastructure?
The way I structure an engineering team depends heavily on the company’s maturity and stage. In an early-stage startup, velocity is the primary currency. At this phase, I align the team so that every member is empowered to make informed, critical decisions autonomously. My role as a leader is to set overarching goals and constantly reiterate priorities. As a company scales toward national deployment, specialized focus areas inevitably emerge. The danger here is the creation of information silos. To prevent this, I employ cross-expertise, goal-centric teams. By grouping engineers from different disciplines around a single objective, we reduce barriers and maintain that nimble decision-making characteristic of a smaller startup. I prioritize building “safety nets” into our infrastructure. This includes robust A/B testing frameworks, staging environments, and highly controlled deployments. When engineers know they have a reliable safety net, they don’t have to make the difficult trade-off between performance and speed.
Looking ahead, as Edge AI continues to evolve across industries like agriculture, retail robotics, and healthcare, what do you see as the most important breakthroughs or shifts needed to make these systems more scalable, efficient, and widely adopted?
We are navigating a fundamental shift in how technology interacts with the physical world, one where computers become embedded into the objects we interact with daily. The next frontier is truly ubiquitous Edge AI.
The most important breakthrough will be the clear identification of societal priorities. We have a real opportunity to use Edge AI to make high-quality healthcare and food safety universally accessible.
In agriculture, edge ML is already helping make farming profitable again. In healthcare, the same computer vision principles I applied to vocal fold kinematics can be scaled to enable non-invasive, continuous monitoring in elderly and pediatric care. To get there, we need to move toward context-aware autonomy and systems that are not just intelligent, but genuinely responsive to the human environment around them. Deployed ethically and efficiently, Edge AI can become an infrastructure that enhances human safety and well-being at a global scale.
