Dr. Andy R. Terrel (he/him)
Austin, Texas Metropolitan Area
6K followers
500+ connections
About
I am a technical executive / product leader / ML Engineer with experience in several…
Activity
6K followers
-
Dr. Andy R. Terrel (he/him) shared thisCome build great things with the community!Dr. Andy R. Terrel (he/him) shared thisI’m hiring! We’re looking for someone to lead NVIDIA’s open source sustainability efforts. If you’re excited about building programs in the age of AI to support open source communities, please use the link’s below.
-
Dr. Andy R. Terrel (he/him) reposted thisDr. Andy R. Terrel (he/him) reposted thisI’m hiring! We’re looking for someone to lead NVIDIA’s open source sustainability efforts. If you’re excited about building programs in the age of AI to support open source communities, please use the link’s below.
-
Dr. Andy R. Terrel (he/him) reposted thisDr. Andy R. Terrel (he/him) reposted thismlpack received $300K from the Sovereign Tech Agency to strengthen open-source ML infrastructure. Over the next 18 months, you can expect: modern GPU backends, ONNX support, better Python/R/Julia bindings, DuckDB + Arrow integration, and more. Read more on the NumFOCUS blog: https://hubs.la/Q04f4Pkk0mlpack receives $300K to support open source AI developmentmlpack receives $300K to support open source AI development
-
Dr. Andy R. Terrel (he/him) shared thisGreat opportunity!!! Jake is THE AMAZING LEADER behind the most used CUDA library, CCCL. Hitch your wagon to this star and hold on tight.Dr. Andy R. Terrel (he/him) shared thisDo you have experience in multi-GPU/distributed programming? Want to come help build the next generation of Thrust/CUB delightful multi-GPU algorithms in CCCL? If so, hit me up, I'm hiring. Also looking for: - CUDA ninjas who care equally about building high quality abstractions as much as performance - People who know both CUDA and enough about compiler internals to be dangerous US Candidates: https://lnkd.in/g2NYFa-E EU Candidates: https://lnkd.in/gDx28KAr
-
Dr. Andy R. Terrel (he/him) shared thisRunning scikit-learn on GPUs—with CUDA support available the day new NVIDIA hardware ships—is now within reach for the open source community. Join experts from NVIDIA and :probabl. and learn how to accelerate scikit-learn on GPUs, and how OSS codes can use CUDA directly with supported APIs and languages from day one. 📅 May 5 📍 GOSIM 2026 | Station F, Paris
-
Dr. Andy R. Terrel (he/him) shared thisRunning scikit-learn on GPUs—with CUDA support available the day new NVIDIA hardware ships—is now within reach for the open source community. Join experts from NVIDIA and :probabl. and learn how to accelerate scikit-learn on GPUs, and how OSS codes can use CUDA directly with supported APIs and languages from day one. 📅 May 5 📍 GOSIM 2026 | Station F, Paris
-
Dr. Andy R. Terrel (he/him) shared thisGive it a listen to see the work we are doing to help make Python packaging work with accelerated computing!Dr. Andy R. Terrel (he/him) shared thisPodcast = Wheel Next + Packaging PEPs NVIDIA - Astral - Quansight on 'Talk Python' https://lnkd.in/gm2JTDsg
-
Dr. Andy R. Terrel (he/him) shared thisCome join the celebration of scikit-learn and performant data science!Dr. Andy R. Terrel (he/him) shared this📣 Speaker announcement: Dr. Andy R. Terrel (he/him), Senior Manager of Compute Products at NVIDIA and Director at NumFOCUS. 🗣️ Python on GPUs: Speeding Up Machine Learning 📍 Own Your Data Science & AI Workshop 🗓️ May 5, 2026 🎟️ Get your FREE ticket with our code: PROBABL https://lnkd.in/enNMr-JK 🥂 :probabl. and Fundamental will host drinks at La Felicità afterwards. RSVP here: https://luma.com/w7ghw3z2 This workshop is brought to you by scikit-learn, :probabl., and Fraunhofer IAIS. We thank the GOSIM Foundation for hosting us, and Fondation d'entreprise Free for supporting scikit-learn and #OpenSource.
-
Dr. Andy R. Terrel (he/him) shared thisThe Power of Constraints: A Small Experiment in Skipping the Engineering TeamThe Power of Constraints: A Small Experiment in Skipping the Engineering Team
-
Dr. Andy R. Terrel (he/him) liked thisDr. Andy R. Terrel (he/him) liked thisAnd that's a wrap! Thanks to our co-organizers, speakers, and all of the attendees at our Own Your Data Science and AI Workshop. We had a blast and we hope you did, too! Special thanks to the following speakers for your great talks: ☕️ Morning session Nicolas Flores Herr: The Long Game: Building Europe’s Sovereign AI Layer Nouamane Tazi: The Smol Training Handbook: Lessons from Training SOTA LLMs Anastasia Stasenko & Yannick Detrois: Synthetic Data for the Commons: Building Open SOTA LLMs with Synthetic Environments Said Taghadouini: LightOnOCR: Pushing the Performance-Efficiency Pareto Frontier of Open OCR Models Björn Plüster & Benedikt Droste: Evaluating Frontier Agents on Economically Valuable Tasks David Cortes & Adel Chaïbi: Extension for scikit-learn: Accelerate Workflows with Two Lines of Code Violetta Mishechkina: From agent traces to analytics Sajid Alam: Kedro: The Toolbox for Production-Ready Data Pipelines ☕️ Afternoon session Gael Varoquaux: TabICLv2: A Better, Faster, Scalable & Open Tabular Foundation Model Stefanie Senger, PhD & Olivier Grisel: The Adoption of the Python Array API in scikit-learn Dr. Andy R. Terrel (he/him): Python on GPUs: Speeding Up Machine Learning Tim Head: Accelerating scikit-learn with GPUs Evgeni Burovski: SciPy's Support for GPUs, PyTorch, and CuPy ✨ Marie Sacksick & Fabien Pesquerel: Skore: Track, Explore, and Share Your Data Science Workflows Sylvain Corlay: notebook.link: Run notebooks instantly in your browser David Arturo Amor Quiroz: Skolar and the Value of Curated ML and AI Skills Education Joanna Kramer & ✨ Marie Sacksick: What Open Source Communities Can Do to Promote Responsible AI Watch this space for our workshop highlights... 👀
-
Dr. Andy R. Terrel (he/him) liked thisDr. Andy R. Terrel (he/him) liked thisYesterday I showed the full CUDA compilation pipeline. One thing the diagram makes clear: nvcc is not a compiler. It's a driver. It orchestrates several compilers, each responsible for one specific transformation. Here's what's actually running under the hood. 𝟭. 𝗴𝗰𝗰 -𝗘 (𝗽𝗿𝗲𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗼𝗿) First pass. Expands macros, inlines headers, resolves #ifdef. The output is a .ii file: plain preprocessed C++, no CUDA extensions yet. This is what every subsequent compiler actually reads. 𝟮. 𝗰𝗶𝗰𝗰 (𝗡𝗩𝗜𝗗𝗜𝗔'𝘀 𝗱𝗲𝘃𝗶𝗰𝗲 𝗳𝗿𝗼𝗻𝘁𝗲𝗻𝗱) This is the one most people don't know exists. cicc takes the .ii file, extracts device code, and compiles it to PTX. It handles everything CUDA-specific: __global__, __shared__, threadIdx, template instantiation for device functions. The output is architecture-independent PTX. cicc is a full optimizing compiler with its own IR, its own analysis passes, its own fusion decisions. 𝟯. 𝗽𝘁𝘅𝗮𝘀 (𝗣𝗧𝗫 𝗮𝘀𝘀𝗲𝗺𝗯𝗹𝗲𝗿 𝗮𝗻𝗱 𝗿𝗲𝗴𝗶𝘀𝘁𝗲𝗿 𝗮𝗹𝗹𝗼𝗰𝗮𝘁𝗼𝗿) ptxas takes PTX and produces SASS, the real ISA specific to one architecture (sm_80, sm_90, sm_120). This is where register allocation happens, where instructions get reordered to hide latency, and where control codes are computed and embedded into every 16-byte instruction slot. The output is a cubin. 𝟰. 𝗴𝗰𝗰 / 𝗰𝗹𝗮𝗻𝗴 / 𝗠𝗦𝗩𝗖 (𝗵𝗼𝘀𝘁 𝗰𝗼𝗺𝗽𝗶𝗹𝗲𝗿) The host code goes to your system compiler, unchanged. nvcc just passes it through with a few injected stubs for kernel launch (<<<>>>). The host compiler has no idea it's part of a GPU build. 𝟱. 𝗴𝗰𝗰 -𝗰 (𝗲𝗺𝗯𝗲𝗱𝗱𝗲𝗿) Once the fatbin is assembled (one or more cubins plus PTX), nvcc uses the host compiler again to embed it as a binary blob inside the host object file. The GPU code literally lives inside your .o as a data section. 𝟲. 𝗴++ (𝗹𝗶𝗻𝗸𝗲𝗿) Final link against libcudart. The runtime library provides the glue: it reads the embedded fatbin at startup, talks to the driver, and handles kernel launch. Why does this matter? Because each compiler makes independent decisions. cicc decides how to lower your C++ to PTX. ptxas decides how to schedule and allocate registers. If your kernel is slow, the bug could live in either one, and they leave different traces. Hope it is clearer now 😶🌫️
-
Dr. Andy R. Terrel (he/him) liked thisDr. Andy R. Terrel (he/him) liked thisToday's our 40th anniversary celebration at National Center for Supercomputing Applications (https://lnkd.in/dk-BeHjA) NCSA was founded in 1986, and I started using NCSA resources (a Cray 2) in 1989 or 1990, I think, and I've been working here for 10 years, since the 30th anniversary.
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
-
University of Chicago
-
-
Specialization in Computational Mathematics, Worked as System Administrator and Teaching Assistant for Introductory Programming Courses
-
-
-
-
-
-
-
-
-
Volunteer Experience
-
-
-
Advisory Board Member
The Iron Yard
- 2 years 4 months
Education
Helped with curriculum review, mock interviews, supported class presentations, participated in quarterly board reviews, and gave guest lectures on negotiating your first job.
-
Managing Director Austin
Founder Dating
- 4 months
Science and Technology
Ran monthly meetings and connected local entrepreneurs to wider FounderDating network.
Publications
-
Finite Element Integration with Quadrature on the GPU
ArXiv
We present a novel, quadrature-based finite element integration method for low-order elements on GPUs, using a pattern we call thread transposition to avoid reductions while vectorizing aggressively.
Other authors -
-
SymPy: Symbolic computing in Python
PeerJ CS Preprints
See publication
SymPy is an open source computer algebra system written in pure Python. It is built with a focus on extensibility and ease of use, through both interactive and programmatic applications. These characteristics have led SymPy to become the standard symbolic library for the scientific Python ecosystem. -
Scientific Software Communities
IEEE Computing in Science and Engineering
-
Finite Element Intergration on GPUs
ACM Transactions on Mathematical Software
See publicationWe present a novel finite element integration method for low-order elements on GPUs. We achieve more than 100GF for element integration on first order discretizations of both the Laplacian and Elasticity operators on an NVIDIA GTX285, which has a nominal single precision peak flop rate of 1 TF/s and bandwidth of 159 GB/s, corresponding to a bandwidth limited peak of 40 GF/s.
-
ForestClaw: Hybrid forest-of-octrees AMR for hyperbolic conservation laws
submitted to International Conference on Parallel Computing - ParCo2013
We present a new hybrid paradigm for parallel adaptive mesh refinement (AMR) that combines the scalability and lightweight architecture of tree-based AMR with the computational efficiency of patch-based solvers for hyperbolic conservation laws. The key idea is to interpret each leaf of the AMR hierarchy as one uniform compute patch in $\sR^d$ with $m^d$ degrees of freedom, where $m$ is customarily between 8 and 32. Thus, computation on each patch can be optimized for speed, while we inherit the…
We present a new hybrid paradigm for parallel adaptive mesh refinement (AMR) that combines the scalability and lightweight architecture of tree-based AMR with the computational efficiency of patch-based solvers for hyperbolic conservation laws. The key idea is to interpret each leaf of the AMR hierarchy as one uniform compute patch in $\sR^d$ with $m^d$ degrees of freedom, where $m$ is customarily between 8 and 32. Thus, computation on each patch can be optimized for speed, while we inherit the flexibility of adaptive meshes. In our work we choose to integrate with the p4est AMR library since it allows us to compose the mesh from multiple mapped octrees and enables the cubed sphere and other nontrivial multiblock geometries. We describe aspects of the parallel implementation and close with scalings for both MPI-only and OpenMP/MPI hybrid runs, where the largest MPI run executes on 16,384 CPU cores.
Other authors -
-
Symbolic Statistics with SymPy
Computing in Science and Engineering
Replacing symbols with random variables makes it possible to naturally add statistical operations to complex physical models. Three examples of symbolic statistical modeling are considered here, using new features from the popular SymPy project.
Other authors -
-
From equations to code: Automated Scientific Computing
Computing in Science and Engineering
See publicationAn example of using code generation to keep the readable high level description of code while still obtaining efficiency of low-level optimizations.
Projects
-
Multiphase Flow in porus media and uncertainity quantification
- Present
Other creators -
-
Memex
-
See projectPI building tools to crawl and analyze deep and dark web. Tools include:
Image analysis: https://github.com/memex-explorer/image_space
Topic analysis: https://github.com/memex-explorer/topic_space
Domain search analytics: https://github.com/memex-explorer/memex-explorer -
Blaze
-
See projectThe Blaze ecosystem is a set of libraries that help users store, describe, query and process data.
-
SciPy2013 Conference
-
A scientific computing conference consisting of two days of tutorials followed by two days of presentations, and concludes with two days of developer sprints on projects of interest to the attendees.
Other creators -
Organizations
-
Python Software Foundation
Managing member
Recommendations received
7 people have recommended Dr. Andy R.
Join now to viewOther similar profiles
Explore more posts
-
Rahul Dharankar
The North Texas AI community continues to lead practical, implementation-focused discussions bridging theory with real-world results. Recent IoT meetup in North Texas brought together technology leaders, manufacturing experts, and AI practitioners for deep-dive discussions on enterprise AI Participants (Partial List): Walker Reynolds - IIoT & Industry 4.0 Solutions Architect Tom Hamilton - Telecom Infrastructure & Inventory Optimization Alex Arias (Texas Instruments) - Smart Market Systems Engineer Nicholas Yarmesch (Uplink) - IoT Deployment Platform Founder Hubert Rhasaj (Oyap) - Service Business Digital Transformation Mohammed Hassan - Energy Systems & Battery Management Developer Carl Daley (Valueyetics) - AI Consultancy Founder Sergey Rossi - Predictive Maintenance Solutions (25+ years experience) Ron Assumption - AI Copilots for Wireless Networks Co-Founder Gavin Lehotsky (Texas Automation Systems) - Plant Automation Specialists Key Demonstrations & Case Studies: Walker showed a live demonstration of: UNS Studio - 6.5M line IIoT platform enabling real-time manufacturing dashboard generation in minutes rather than weeks Model, Context, Protocol (MCP) framework showcase - demonstrating AI agents interacting with existing enterprise systems Bible application development case study - 160K lines of code completed in 4 days with Cursor The discussions emphasized three critical themes: Data Revolution: Moving from unreliable manual data collection to AI-supervised precision Speed-to-Value: Dramatic reduction in development cycles through agentic AI implementation Human-AI Synergy: Leveraging AI for execution while preserving human advantages in creativity, empathy, and strategic thinking Key Takeaways: Data fidelity breakthrough: AI agents achieve 99.999% accuracy when human-supervised vs. 98% error rate in traditional collection - using agentic AI - achieving 99.999% data fidelity vs. traditional 2% human accuracy (98% "pencil-whipped") Development acceleration: Complex applications now deployable in days rather than months through agentic frameworks Industry resistance reality: Manufacturing adoption remains slow despite proven ROI, citing trust and cultural barriers Employment evolution: Focus shifting from replacement fears to skill augmentation and new role creation Energy implications: Nuclear power sentiment shifting positive among environmentalists as AI demands increase Human differentiation: Emotion, intrinsic motivation, and empathy remain uniquely human competitive advantages These session highlighted that we're entering an unprecedented era where AI could create true abundance for the first time in human history, while emphasizing that success depends on thoughtful human-AI collaboration rather than wholesale automation. #AI #Manufacturing #IoT #AgenticAI #DigitalTransformation #DataFidelity #TechLeadership #Innovation
2 Comments -
Khaled Zaky
Anthropic shared a wild case study on Claude Opus 4.6. When the model hit a wall during a websearch test, it didn't just fail as it correctly hypothesized it was being evaluated, identified the specific benchmark, and then wrote its own code to find and decrypt the answer key 🤯 As I’ve written recently on my blog, the shift from "software that executes" to "agents that act" changes everything. This is a perfect example of why governance is an architectural problem, not a compliance one. When an agent is smart enough to "hack the test" to achieve its goal, traditional static gates and simple prompts aren't enough. We need a true platform mindset…one built for autonomous actors that can recognize the boundaries of their sandbox and actively look for a way out….If you’re still treating AI as a deterministic tool, it’s time to rethink your stack The full deep dive from Anthropic is worth a read: https://lnkd.in/gPXdffpy
5 Comments -
Pierluigi Failla, PhD
mHC: Is the Residual Connection obsolete? 🧠 ⏲️ For a decade, ResNet has been the backbone of Deep Learning. But as models scale, the standard "identity mapping" is becoming a bottleneck. DeepSeek-AI's new paper on Manifold-Constrained Hyper-Connections (mHC) proposes a fascinating shift. By projecting residual streams onto the Birkhoff polytope (doubly stochastic matrices), they’ve managed to expand bandwidth without the exploding gradients that usually follow. If you want to take a look on how this changes the macro-design of LLMs check this. 👇 https://lnkd.in/dhkWZD-V #AI #DeepLearning #DeepSeek #NeuralArchitecture #LLM
-
WebCraftingCode
In the latest episode of AXRP, Lee Sharkey introduces Attribution-based Parameter Decomposition (APD), a groundbreaking approach to enhancing interpretability in deep learning models. APD shifts the focus from traditional activation-based methods to parameter-centric analysis, offering a three-term optimization framework that ensures faithfulness, minimality, and simplicity in model interpretation. By decomposing weight parameters rather than activations, APD reveals latent computational mechanisms with greater clarity. The implications of this work are significant. As AI systems grow in complexity, understanding their inner workings becomes essential for building trust and ensuring robustness. APD not only provides a pathway for clearer interpretability but also enhances the capability of models to discover and utilize sparse, low-rank structures efficiently across various architectures. This innovative approach positions APD as a vital tool for researchers and practitioners aiming to navigate the challenges of AI transparency. Learn more about this exciting development in AI interpretability [here](https://lnkd.in/eQKcRJxn).
-
Sravan Bodapati
Anthropic's recent blogpost discusses the important shift from evaluating simple LLMs ---> testing complex AI agents. It highlights why standard benchmarks fall short for agentic workflows and offers a practical roadmap for building robust, multi-turn evaluations." Packed with insights, but couple of takeaways that stood out to me: 🎯 1. 𝐂𝐨𝐦𝐩𝐨𝐮𝐧𝐝𝐢𝐧𝐠 𝐯𝐚𝐥𝐮𝐞 𝐨𝐟 𝐭𝐡𝐞𝐬𝐞 𝐬𝐲𝐬𝐭𝐞𝐦𝐬: where costs are visible upfront but their benefits accumulate over time. (quicker iterations, faster deployment) 🎯 𝟐. 𝐓𝐡𝐞 𝐂𝐨𝐧𝐬𝐢𝐬𝐭𝐞𝐧𝐜𝐲 𝐅𝐢𝐱 / 𝐃𝐞𝐭𝐞𝐫𝐦𝐢𝐧𝐢𝐬𝐦 𝐢𝐧 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧𝐬: Regardless of agent type, agent behavior varies between runs. Sometimes, we want to measure how often (what proportion of the trials) an agent succeeds for a task, while for others, we want to measure if Agent succeeds atleast once. 🚀 Single success is enough across runs ? Use pass@k. 🚨 Need reliability across runs? Use pass^k. The post contains many more valuable insights worth exploring. Read more here: https://lnkd.in/gwCxqhDj
2 Comments -
Shane Deen
Claude just released a memory import tool: https://lnkd.in/eM8c7Gf5 You paste a prompt into ChatGPT (or whatever you're using), it spits out your context — preferences, projects, how you work — and you drop it into Claude's settings. That's it. Claude picks it up from there. The switching cost has always been the real barrier. Not the AI itself, just the fact that you've spent months building context somewhere else. This is Anthropic admitting that out loud and actually doing something about it. Quiet release. Worth knowing about.
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content