Open to opportunities

Poojith Devan

@poojithdevan

Message

I build benchmarked LLM inference stacks with Triton kernels, quantization, and vLLM.

India

Message

What I'm looking for

I'm looking for an ML systems or inference engineering role where I can build, optimize, and rigorously benchmark efficient LLM serving and quantization systems, ideally in a remote role available immediately.

At OneBit, I built packed ternary Triton inference kernels that increased Qwen3-8B decode from 1.48 to 15.5 tok/s while reducing VRAM from 15.3 to 7.4 GiB on an RTX 5070. I benchmark latency, throughput, VRAM, and quality rigorously, including 100% argmax agreement against reference outputs.

I've also implemented integer-only Q1.15 inference for FPU-less embedded hardware at Trebuchet System, reaching 77.71% CIFAR-10 accuracy on ResNet-20. My projects span vLLM serving benchmarks, quantization tradeoffs, and OpenAI-compatible LLM gateways with routing, semantic caching, failover, and rate limiting.

Experience

Work history, roles, and key accomplishments

Current

Research Intern

Current

OneBit

Apr 2026 - Present (5 months)

Built a Triton kernel for packed ternary weights, improving Qwen3-8B decode speed from 1.48 to 15.5 tok/s and reducing VRAM from 15.3 to 7.4 GiB on a 12 GB RTX 5070. Verified 100% argmax agreement with reference and profiled with Nsight Compute to optimize memory access.

Triton CUDA VLLM Quantization PyTorch Python Benchmarking