Less Wright

High speed Inference architecture and custom GPU Kernels.

PyTorch Distributed: Distributed Training - core contributor, TorchTitan which showcases 1D-5D parallelism, with FSDP2, CP, ZB/ZBV PP, asyncTP, fp8. https://github.com/pytorch/torchtitan Distributed Inference - core contributor, TorchChat Distributed Inference showcasing PP/TP/SP and soon asyncTP: https://github.com/pytorch/torchchat PyTorch Distributed Compiler, Graph Optimizations Cuda Kernels - Fused, Paged and Quantized AdamW, Stochastic rounding (BF16, FP16) Triton Kernels - FlashAlibi, RMSNorm, SplitK and StreamK matmul, groupGEMM in progress. Quantized Optimizers (4bit) Accelerating Transformers with Parallel Attention Blocks PyTorch Tensor Parallel, 2D and 3D (Pipeline Parallel) Speaker, PyTorch Conference 2024 "Efficient Scaling of Pre-Training LLMs with TorchTitan and AWS Sagemaker" Speaker, PyTorch Conference 2023 PyTorch FSDP (Fully Sharded Data Parallel) distributed training for AI * AnyPrecision Bfloat16 optimizer with Kahan summation * Presenter at Nvidia Fall GTC 2022, SuperComputing 22 * PyTorch YouTube video series on FSDP1 * Collaborative efforts with IBM Research, AWS

Cloud-based AI architecture and pipeline development for diagnostic detection and classification of infectious diseases, with scaling up to country level deployment. Covid-19 (antigen, antibody, genetic PCR), Malaria, HIV, Influenza A/B. * Core AI Architecture now US patented - US10956810B1 * AI exceeds human expert detection in accuracy. Object detection via multi-chain Transformer Architectures (DETR, SOLQ) and classification (ResNeST, EfficientNet, MixNet, resMLP_Mixer) and VIsual Transformers including Nest and CS-Win Transformer. Developed multiple image augmentations to provide enhanced visual diagnostics, became core algo for additional product offering. Synthetic Image Generation via Unity Perception, GANs and unpaired image translation (CUT, sinCUT, StyleGan2, DCGAN). FB AI Research github contributor (DETR - Detection Transformer). Mobile based image segmentation (binary) and mobile classification. No Reference Image Quality Assessment - Blur detection, glare detection, low contrast and shadow detection (w Pooling Image Transformer)

We are a robotics company that produces lab grown gems, including Amora, the worlds most stunning gem, via our TIGER robots. Inc5000 2016 list of fastest growing private companies. #2494 overall, and #64th fastest growing manufacturer in the US.

AdaPools.io (http://AdaPools.io) is a prototype global staking pool and smart node (AI monitored) network for the Cardano / Ada cryptocurrency. Staking for Ada is still in development, but will go live in 2Q 18, and we have a fully functioning global network running Cardano SL (and even Explorer SL) nodes in preparation. This project was an outgrowth of the expansion for BetterThanDiamond to be able to accept payment in cryptocurrency and will spin off into it's own entity in the future.

Designed user friendly block-chain explorer site for Cardano (Ada) cryptocurrency and extended it to Ethereum. The main goal for this was to make it easy for users to verify an Ada address before sending, as well as to monitor transactions in and out of wallets. We then added in the same functionality for Ethereum, with Adacheck.io dynamically adjusting between blockchains to allow single use for multiple currency wallets.

Founded Seattle and Eastside (East Seattle area) Meetup for Pytorch and FastAI Deep Learning AI users. 30+ members.

Co-wrote the MultiCam software program (C++) to monitor and record in MP4 video, supporting up to 16 wireless cameras. Added motion detection, timestamps, and emailing of photos to allow users to be alerted to visitors, deliveries, etc. Program was initially designed to be a freebie provided along with each wireless camera, but proved so successful that they charged $100/copy and became highest grossing product for company (~ $10K/day in orders). Articles: Wrote this article on codeproject detailing how the UI was done for this software using WTL and DDX/DDV to handle it efficiently - 5/5 rating from 36 users.

Worked as both perm (2 times - 1994, 1998-2000) and as contractor via my own company which became BetterThanDiamond now. Primarily Microsoft Access from 1.1 ->2000, Office and Visual Data, along with an initial start in MS-DOS.

Online via Coursera - Algorithms: Divide and Conquer, Sorting and Searching and Randomized Algorithms.

MOOC course of their actual univ course "Elements of AI". No programming, but general overview of AI both technically (CNN, back propagation, etc) and in terms of society impact. In theory they will email a certificate but for now I just clipped the congrats at end of course screen.

C Sharp (C#)

Open CourseWare - Solid State Chemistry

U of Michigan Python Specialization. This is the Python Basics, 1 of 5 for their specialization.

2 of 5 for "Python 3 Programming Specialization" via Coursera