Machine Learning Engineer | Systems, Hardware & AI
Imperial College London (EIE)
I like building things that sit close to the hardware and making them fast. Most of what I do ends up somewhere between chip design and machine learning: I'm into how GPUs and accelerators are actually built, how models get squeezed onto them efficiently, and what it takes to get inference running quickly at scale. I spent a few months at Canopy Labs doing exactly that, optimising inference for large video models on H100s.
Outside of that, I just enjoy building systems from scratch to understand them properly, whether that's a CPU, a shell, or a trading engine. If I can't explain what's happening a layer below what I'm working on, I go find out.
- GPU inference optimisation and getting large models to run fast
- Computer architecture and hardware acceleration (FPGA, RISC-V)
- Machine Learning and how you can use it to complete complex tasks
- Experimenting with hardware from time to time
PPO agent deployed on a physical smart-grid testbed
- Custom Gymnasium simulation of a microgrid (battery storage, PV generation, stochastic pricing)
- Deployed the trained policy to real hardware (Raspberry Pi Pico) with live Flask telemetry
Real-time trading system on dual PYNQ-Z1 FPGAs
- FPGA-accelerated linear regression training via incremental outer-product accumulation (~68x faster weight updates)
- Live Binance market data ingestion with a custom microstructure feature pipeline and 512-bit AXI datapath
- 57.6% directional accuracy, 112% paper-trading return under leveraged simulation
5-stage pipelined CPU with cache (SystemVerilog)
- IF/ID/EX/MEM/WB pipeline with hazard detection and data forwarding
- 2-way set-associative write-back cache with LRU replacement and a miss-handling FSM
Custom Unix-like shell written in C
- Recursive, parentheses-aware parser for nested subshells and pipelines
- Process management via
fork,exec,pipe,dup2, andwait
Python, C++, C, SystemVerilog, SQL, Bash, Lua
PyTorch, TensorFlow, Transformers, NumPy, Reinforcement Learning (PPO), GPU inference optimisation (FlashAttention, torch.compile, BF16, KV-cache)
FPGA, RISC-V, Computer Architecture, Multithreading, WebSockets, Unix/Linux
Docker, Terraform, AWS, GCP, PostgreSQL, Git
Flask, FastAPI, REST APIs, MongoDB, Pandas, OpenCV, MediaPipe
- Website: https://sumukhadiraju.com
- LinkedIn: https://linkedin.com/in/sumukh-adiraju-07235b31b
- Email: sadiraju06@gmail.com