Hi, I’m Aayush. I like hard problems, and I like finishing them properly.
I’m studying Computer Science at SRM Institute of Science and Technology, and I’ve spent about a year and a half building software at two companies.
I don’t box myself into one stack or one layer. I learn what the problem needs, build it carefully and stay with it until it holds up in the real world.
A compute platform that gathers underused GPUs from data centres, college labs and workstations, and turns them into production inference endpoints. Customers deploy open-source models; hosts earn from capacity that would otherwise sit idle.
Press a scenario to send it through the system, or select any part to read what it does.
Eight Go services for the control and inference planes: gateway, control API, scheduler, coordinator, trust engine, billing meter, inference gateway and request router.
A Rust host agent that detects and benchmarks GPUs, opens outbound-only WireGuard tunnels and runs each job in its own Firecracker microVM after verifying a signed manifest. It idles under 2% CPU and 300 MB of memory.
Hardware benchmarks and fingerprints feed a 0–100 reputation score for every host.
Customer and host consoles in Next.js with 20+ modules, a 4-step deploy wizard and live dashboards. The platform is designed to grow from 9 to 15,000+ GPUs.
MAX
A local agent that proves it did the job.
2026
An autonomous operator for macOS that you control by voice or text. It reads the live desktop through accessibility trees, plans with a local model and only reports success when it can prove the result on the machine. Nothing leaves the laptop.
MAX's closed loop, illustrated with example requests
Pick a scenario
Press a scenario to send it through the system, or select any part to read what it does.
Targets controls by their accessibility role and label instead of screen coordinates, and refuses to act when two candidates are equally likely.
Separates running a step from achieving the goal: every outcome is proven, unknown or failed, based on process, file, UI and terminal state.
Classifies actions into five risk tiers, hard-blocks destructive commands and writes a hash-chained audit log.
Runs fully offline with Ollama models and Whisper speech recognition on Apple Silicon, covered by 228 passing tests.
SentinelGIS
Early warning for disease outbreaks across India.
2026
A surveillance and forecasting platform for public-health teams. It turns government outbreak reports and live news into district-level risk forecasts and alerts on an interactive map.