Artificial intelligence is moving out of large data centres and into phones, earbuds, doorbells and factory sensors, many of them running on small batteries. In an exclusive conversation with The Free Press Journal, Nikunj Kotecha, a California-based AI and Machine Learning Engineer who works on brain-inspired AI chips, talks about dividing AI work between processors, making AI work in Indian languages, teaching students how AI runs on real hardware, the impact of his contributions to free open-source AI software and the personal assistant he plans to build next.
Nikunj, your career began as a research assistant in the Machine Intelligence Laboratory at the Rochester Institute of Technology (RIT), using AI to translate sign language, and later moved to AI chips at BrainChip. How did that happen, and what do you focus on today?
I did my Master's in Computer Science at RIT in New York, specializing in Artificial Intelligence, years before ChatGPT made AI an everyday word. In the lab, I built deep learning models that translate American Sign Language videos into English text, and I published this research with IEEE and ACM in 2021. That year, I joined BrainChip, which makes Neuromorphic AI chips that are designed to work a little like the brain. With my work there, I showed that a good model also has to fit the device it runs on. In August 2025, I joined Coastaloupe, an AI engineering solutions company in California, as a Senior Machine Learning Engineer. Through Coastaloupe, I also work with BrainChip as a consultant on its chips and software. Today, my focus is getting AI to run well on small devices instead of in the cloud.
Recently you wrote an article for BrainChip called "Two AI Brains Are Better Than One". What do you mean by two AI brains?
Think of a smart doorbell, a pair of earbuds or a sensor on a factory motor. Most of the time, nothing happens, yet a big processor stays awake all day, using battery power just to confirm that nothing has happened. I believe most smart devices should have two AI brains: one that senses and one that thinks. A small, low-power chip watches the sensors all the time and wakes a bigger processor only when something matters. Engineers call this heterogeneous computing. It simply means giving each part of the work to the processor that suits it.
Where is this approach already in use, and how can a chip like BrainChip's Akida bring it to other everyday devices?
Your phone already works this way. Recent phone chips have a large AI engine for jobs such as editing photos and a tiny one that senses all day. Most other devices are built around a processor someone else made, so they need a separate low-power AI chip next to it. That is where Akida fits. It saves power because it skips zeros. Inside a typical AI model, about half of the values passed along at any moment are zero, and with the right training, more than 70 percent can be. I wrote about this idea in BrainChip's technical paper, Activation Sparsity: Unlocking Efficient Deep Learning at Scale.
In August 2026, BrainChip released the Symphony Community Akida Bundle, free software that brings its Akida chips into IBM Spectrum Symphony. You led the release and wrote most of the bundle's code. Why does that matter to businesses?
The same idea works at a much bigger scale. Banks and other large companies have used IBM Spectrum Symphony for many years to spread millions of calculations across shared computers. Small AI jobs in these systems usually go to a GPU, even when a much smaller chip could do them with less power. The bundle builds on earlier integration work at IBM and lets Symphony manage Akida chips just like the usual CPUs and GPUs, so small AI jobs can go to them. The code is open source, on GitHub as symphony-akida under the Apache 2.0 licence, so any developer can download it and adapt it. I built all three of its public demonstrations and authored its technical whitepaper, BrainChip Akida as a Managed Resource in IBM Spectrum Symphony. I also explained the work in a BrainChip post, Multiple Neuromorphic Chips, One Grid, No New Middleware.
Leading AI labs now rely on outside experts to make their models more reliable before people use them. You are a fellow in Handshake AI's MOVE program and one of its reviewers. What does the work involve, and what difference does it make?
A model can look impressive in a demonstration and still give wrong answers to the questions people actually ask. MOVE, short for Model Validation Expert, connects experts with AI labs to improve their models. We write difficult problems, review the models' answers and explain where their reasoning goes wrong. Handshake says all eight of the largest frontier AI labs train with its data, so the mistakes we find help these labs fix weak spots in their models. In September 2025, the project team recognized me as one of its top experts and asked me to record walkthroughs of my approved tasks to train new fellows. A few weeks later, they made me a reviewer in the project's first review layer, where I check other fellows' work.
For Indian readers, one of your most relevant projects is IndicMMLU-Pro, a benchmark that tests AI models in Indian languages, and you are an author of its research paper. What did it reveal about today's AI models?
More than 1.5 billion people speak Indian languages, yet AI language models are mostly measured in English. IndicMMLU-Pro takes a demanding English benchmark of knowledge and reasoning and brings it into nine Indian languages, including Hindi, Marathi, Gujarati, Bengali and Tamil. In 2024, we put the dataset on Hugging Face, where AI researchers share their data, so anyone can download it for free. It has more than 12,000 questions in each language, nearly 109,000 in all. Hugging Face counts more than 7,600 downloads so far. The paper was then published in July 2026 by the Association for Computational Linguistics. The strongest model we evaluated at the time, GPT-4o, answered only about 38 to 45 percent of the questions correctly, depending on the language.
In 2022, you founded BrainChip's University AI Accelerator Program, and now you also mentor working professionals through Great Learning. What do you want students and engineers to take away?
The program started with a pilot in a Carnegie Mellon University course. Since then, I have expanded it to Arizona State, RIT, Cornell Tech, Penn State and other Tier 1 universities in the US. We have now also recorded these sessions, and new universities can sign up for access. At Great Learning, I mentor professionals in its online AI programs, such as the Post Graduate Program in Artificial Intelligence and Machine Learning: Business Applications from the McCombs School of Business at The University of Texas at Austin. I advise all of them: AI is not only about the model anymore. You need to understand the hardware it runs on and how to check that it really works.
Last year, GitHub reported that India had overtaken the United States as the largest community of open-source contributors on its platform, and six of its ten fastest-growing projects were AI tools. You contribute to two tools for AI coding agents, gh-axi and pi-code. Why do you take part, and how does this work help AI development?
When an open-source tool gets better, everyone who uses it benefits. Much of AI is now built this way: GitHub's report also counted more than a million public projects that use large language models. Many developers write that code with AI coding agents, and I contribute to tools that make these agents better. One is gh-axi, which helps these AI agents work on GitHub with short, structured replies, so they finish tasks faster and at lower cost. gh-axi has been downloaded more than 85,000 times since it came out in March. I have also contributed to pi-code, which helps developers who use pi, an open-source coding agent that works with hundreds of AI models, including ones that run locally on their own computer. pi-code reads the setup they already made for Claude Code, another AI coding agent, such as their rules, commands and skills, and adds features pi lacks, such as to-do lists, checkpoints to undo changes, memory and web search. That way, developers can work in pi without setting everything up again. pi-code has been downloaded more than 26,000 times since it came out in July.
You are a Senior Machine Learning Engineer at Coastaloupe, a California AI engineering company, where you will lead the work on a personal AI assistant that runs on the device itself. What will it do, and where do you want to take this work?
So far at Coastaloupe, I have mostly done client work for BrainChip, such as leading the work on AkidaTag. The assistant will be our own project. It will help with the small, quick tasks of a normal day: a student planning the week, a parent making a list for the day's errands between school runs, a young person setting reminders for bills at a first job, or an older person who needs a reminder to take medicines. It will run on the phone itself, so it answers quickly, works without the internet and keeps personal details on the device. We will compare it with a cloud assistant on speed, energy use, how well it does each task and what data leaves the phone, and share a test package so other engineers can check our results.
