skip to content
arjun nair

projects

Things I've built and a couple I'm still working on.

01in progress

Mosaic - Programmable Neural Processing Unit

8-bit NPU · SystemVerilog + Python

Mosaic is a custom 8-bit neural processing unit inspired by NVIDIA's open-source NVDLA architecture. I'm designing a small neural network accelerator from the ground up in SystemVerilog, with Python reference models to check correctness. The proposed design uses a 2×2 multiply-accumulate array and local buffers for weights and activations. I'll compare on-chip reuse and intermediate-result forwarding against extra memory transfers by measuring execution cycles, memory accesses, and data movement.

02completed

Adjustable AC/DC Power Converter

Circuits class · Stanford

I designed, soldered, and characterized an adjustable AC/DC power supply for my Circuits class at Stanford, progressing from full-wave rectification and capacitive filtering to Zener regulation and a MOSFET source-follower output buffer. The finished circuit adjusts from roughly 0–20 V and holds its output within a few hundred millivolts as the load goes from 10 kΩ down to 100 Ω.

03completed

bare-metal motor controller

ARM Cortex-M3

I built a bare-metal motor-control system around a real-time PID loop, working directly with the hardware instead of relying on an operating system to hide timing and peripheral behavior. The project combined feedback control with deterministic embedded software and direct hardware interaction.

04completed

distributed key-value store

C++

I built a replicated key-value store in C++ to explore what happens when networking, storage, concurrency, replication, and failures all become part of the same system. Rather than stopping at a basic GET/SET dictionary, I worked on client/server networking, replicated state, leader/follower behavior, failure recovery, and persistent-storage concepts, then benchmarked the system under load.

05completed

HTTP server

HTTP/1.1

I built a high-performance HTTP/1.1 server to understand what happens underneath a basic web request — sockets, request parsing, concurrency, memory behavior, and performance bottlenecks. Once it handled requests correctly, I benchmarked the implementation and optimized the slow parts.

arjun nair · stanford