Co-founder, G3 AI Labs

Adharsh
Thomas.

I build AI around the limits of real hardware.

My research spans brain-inspired architectures, efficient inference on edge devices, and computer vision.

Current work

G3 AI Labs

We’re building brain-inspired computing systems for AI that runs close to where data is generated.

Sparse computationLocal inference
Visit g3lab.tech

Selected research

Experiments and prototypes across model design, video inference, and image restoration.

Computer vision · Experiment

Hybrid image super-resolution

I tested whether an adversarial model could recover detail lost when CIFAR-10 images were downsampled by 2×. The generator combines multiscale features, residual-in-residual dense blocks, and spatial and channel attention.

Dataset
CIFAR-10
Input to output
16×16 to 32×32
Training run
10 epochs
Two examples from the experiment, a boat and a frog. Each row shows the low-resolution input, the model reconstruction, and the original CIFAR-10 image.
Two qualitative examples from the CIFAR-10 experiment. The boat and frog are shown at the same display size for comparison.

Edge AI · Prototype

Adaptive video inference

Running a detector on every frame spends compute even when a scene barely changes. I’m prototyping a scheduler that uses recent frames and caps how long a region can go without a fresh detection.

What I’m measuring

Detector calls, end-to-end latency, and missed detections against a detector-on-every-frame baseline.

Model architecture · Early research

Brain-inspired architectures

I’m exploring how parts of a model can specialize, how a system can retain context over time, and when it can activate only the computation a task needs.

Public workspace

Brain_Engine is an early workspace for experiments with brain-inspired architectures and Transformers.

Browse Brain_Engine

How I test ideas

What I look for when an architecture has to work within a shared compute, memory, or power budget.

Start with a baseline

For video, that means running the same detector on every frame. For image reconstruction, a simple resize provides a reference for what the learned model adds.

Change one part of the system

Keep the detector fixed when testing a scheduler. In a model experiment, compare against a simpler variant to see what attention or another added component contributes.

Measure the trade-off

Track latency and detector calls alongside missed detections. Inspect hard cases such as motion and scene changes, where reusing an older prediction can cause errors.