Open to new opportunities

Sami Koneru-Ansari

Building ML systems from the ground up.

EECS at UC Berkeley. Currently at Intel building deep learning primitives in oneDNN. Previously optimized LLM inference at Piris Labs, with CUDA kernels and from-scratch paper reimplementations on the side.

GitHub LinkedIn Email Resume

PythonC++CUDAPyTorchSGLangRustTypeScriptReactNode.jsPostgreSQLLangChainDockerKubernetesGCPAWS

Education

University of California, Berkeley

B.S. Electrical Engineering and Computer Science · 4.0 GPA

Relevant coursework

Data Structures and Algorithms
Machine Learning
Machine Structures
Signal Processing
Computer Networks
Discrete Mathematics
Probability Theory
Linear Algebra

Work experience

AI Software Engineer Intern

Intel · Folsom, CA

Sept 2026 — Present

Developing deep learning primitives in oneDNN, Intel's open-source library behind PyTorch and TensorFlow on Intel hardware, working in C++ on Linux across floating-point arithmetic, numerical accuracy, and library architecture.

C++oneDNNLinux

Machine Learning Engineer Intern

Piris Labs · Remote

June 2026 — Aug 2026

Built ML infrastructure to optimize LLM inference on B200 GPUs, spanning automated GCP deployment, SGLang serving, prefix caching, multi-tenancy, and per-instance routing, as groundwork for custom photonics hardware.
Researched and implemented scheduling and batching policies that cut median end-to-end latency by over 50%.
Benchmarked speculative decoding, batching, scheduling, and routing on unified and disaggregated systems, reaching 483 tok/s on GLM 5.2, 23% more than Databricks.

SGLangCUDAGCPKubernetes

Founding Software Engineer

DreamIt · Berkeley, CA

Feb 2026 — May 2026

Directed a 3-person engineering team shipping end-to-end Rust features for an early-stage no-code gaming app, scaling to 1,400+ users.
Architected an async Claude API pipeline that generates playable games from natural-language prompts.

RustClaude API

Data Science Intern

TokenWorks · Remote

Feb 2026 — May 2026

Fine-tuned BART-base as a sequence-to-sequence tagger for low-resource Sumerian, jointly predicting part of speech, lemma, and English gloss from 217K CoNLL-U tokens of transliterated cuneiform in a single decoding pass.
Built a self-hosted Wikibase lexical knowledge graph with idempotent Python importers mirroring Wikidata's Sumerian lexeme schema through a resumable, checkpointed pipeline.

PyTorchBARTDockerWikibase

Machine Learning Engineer (Contract)

GetGreen · Remote

Sept 2025 — Dec 2025

Engineered a RAG chatbot using LangChain and ChromaDB for semantic search over 1,000+ scraped articles, answering sustainability questions in under 3 seconds.
Scraped and processed the retrieval corpus with BeautifulSoup, pandas, and SQL.

LangChainChromaDBpandasSQL

Software Development Intern

California Department of Technology · Rancho Cordova, CA

May 2024 — July 2024

Built full-stack web apps with HTML/CSS/JS for the frontend, Node.js for the backend, PostgreSQL for relational data, and JWT authentication to manage scheduling and file sharing for 100+ employees.
Managed datasets of 500+ employees and 70+ clients using SQL and Excel to drive operational reporting.

Node.jsPostgreSQLJWTSQL

Featured projects

View all projects

CUDA CNN

Machine Learning

CNN built from scratch in C++/CUDA with hand-written backprop, im2col, and GEMM kernels. No deep-learning or matmul libraries. 99%+ on MNIST at ~50% the throughput of an equivalent PyTorch model on T4 GPUs.

C++CUDA

GPT-2 from scratch

Machine Learning

124M-parameter decoder-only transformer with causal self-attention and learned positional encodings, following “Attention Is All You Need.” Optimized CUDA training for an 11× speedup over the unoptimized baseline.

PyTorchCUDAPython