About
I work on how large models route computation internally, and on voice systems that have to answer while someone is still talking. The two turn out to be the same problem viewed from different ends.
Current work
Most of my time goes to foundation model architecture at AntEngage, where I am Founder & CTO. We are building models from scratch — architectures designed from first principles rather than inherited from the transformer lineage. The thesis is straightforward even if the work is not: the efficiency and capability ceiling everyone is currently pushing against is a property of the architecture, not of what is achievable. Something built for the next two decades should not be a derivative of something designed for the last one.
I am not writing publicly about the architecture itself yet. What I can say is that it is not a transformer variant, that efficiency is a design constraint rather than an optimisation applied afterwards, and that the target is capability per unit of compute rather than capability per unit of scale.
At RoboDIB, my hardware startup, I build pcbeditor.com — describe a circuit in plain English and get back a verified schematic and a fabricable board. It is AI-native across the whole flow rather than at one step: schematic, circuit, footprints, netlist and trace routing are generated, design-rule checking and copper pour are done by the model, and the result is rendered as a 3D layer stack and board. Edit by chat, export to KiCad or Gerber. It runs in the browser and it is free. Circuit synthesis and model architecture are closer relatives than they look: both are search over a structured design space where most candidates are invalid for reasons you can check.
Published work
The research on this site is the part I can show. Mixture-of-experts routing and checkpoint compression produced MoE-Watcher-Modifier and a result that contradicts the assumption most pruning work starts from — routing is far flatter than the collapse narrative predicts. Understanding how existing architectures actually allocate their capacity is not a detour from building a new one; it is most of the argument for doing so.
Earlier work on speech runs alongside it. Real-time voice is a brutal constraint: latency is a product feature, streaming architecture is not optional, and a model that is right on average but slow is simply wrong.
How I got here
I started building for other people in 2015, freelancing through Pyloom Innovations. In 2021 I spent a year at Avanze between clients and engineers — the most useful detour of my career, because it taught me that most failed software fails at the translation layer rather than the code.
From 2022 to 2024 I was at GeekyAnts, moving from Software Engineer III to Senior Software Engineer III. Enterprise Java and Spring Boot, microservice decomposition, and a lot of performance work — the kind of engineering where the win is a page that loads faster and nobody notices, which is exactly the point. I've been at AntEngage since April 2024.
How I work
I try not to publish a claim I cannot reproduce. Where a number appears on this site it comes from a run whose command and output are committed somewhere public, and the figures are the ones the tool actually generated rather than redrawn for presentation. When I don't know something I would rather leave the gap visible.
I also give talks on the JVM and distributed systems, and write about engagement systems in healthcare — which is where a lot of the voice work meets an actual clinic.
Elsewhere
Cricket, guitar, long drives. I take consulting work on architecture, ML systems and reviews, and the fastest way to reach me is email.