Voxel51

• Explore, tinker with, and push the boundaries of the latest AI and machine learning technologies to showcase what's possible with open-source tools and frameworks

• Build inspiring prototypes, proof-of-concepts, and example applications that demonstrate creative and impactful uses of AI to the developer community

• Contribute meaningfully to the open source ecosystem by creating and maintaining libraries, tools, datasets, tutorials and guides that enable developers to more easily leverage AI capabilities

• Curate, clean, annotate and publish unique multimodal datasets spanning text, images, audio, video, 3D and other emerging data types to facilitate novel machine learning research and applications

• Create engaging technical blog posts, social media content, videos, and presentations that educate and inspire developers and data scientists to build with AI

•  Collaborate openly with our internal teams and the broader AI/ML community to cross-pollinate ideas, share knowledge, and collectively advance the state-of-the-art

• Evangelize our technology and vision by representing the company at hackathons, conferences, meetups, universities, and other developer-focused events

•  Influence the product roadmap by providing input on the developer experience, uncovering opportunities for new features and improvements based on community feedback and your work

• Continuously learn, experiment with, and provide feedback on beta versions of our latest models and capabilities to help shape the future of our platform

Hacker-in-Residence

LinkedIn

1) Introductory computer vision course for LinkedIn Learning

2) Prompt Engineering with LangChain (releasing March 2024)

3) RetrievalAugemnted Generation with LlamaIndex (releasing Q2 2024)

4) Fine-tuning LLMs with Hugging Face TRL (est release Q3 2024)

5) Text Generation with Hugging Face Transformers (est release Q3 2024)

LinkedIn Learning Instructor

The Artists of Data Science

The ONLY personal growth and development podcast for Data Scientists

Host

Deci AI

Developer Relations Manager

Pachyderm Inc.

Senior Developer Advocate

Comet

ML Developer Advocate

Data Science Dream Job

Principal Data Science Educator

Price Industries

Lead Data Scientist

Bold Commerce

Senior Data Scientist

Emergent BioSolutions

Biostatistician

I create content about deep learning | The DevRel and Deep Learning Dude

Hacking on LLMs, curating multimodal datasets, and writing a book about RAG.

Harpreet Sahota 🥑

sam3 is seriously a step change improvement over sam2

the model is seriously good, glad they brought back text prompts. the automatic segmentation isn't as good as i hoped it to be...but it still works quite well. 

i also got it to output image embeddings and that was cool to visualize in the app

learn more here: https://lnkd.in/gzqJzj6G

🤖 Hacker-in-Residence @ Voxel51| 👨🏽‍💻 AI/ML Engineer | 👷🏽‍♀️ Technical Developer Advocate | Learn. Do. Write. Teach. Repeat.

qwen3vl is literally the best model i've used for video usecases. blows my mind

hosting a virtual event tomorrow about document ai

you may or may not find it interesting, but i guarantee you'll learn something

join me here: https://lnkd.in/dFXPimVc

sony ai released a pretty cool dataset called the fairness human centric image benchmark, super high quality labels

instructions for downloading and parsing the dataset into fiftyone format can be found here: https://lnkd.in/g_VwxJx4

physical AI is top of mind as the next frontier for deep learning. but it'll only succeed if you have good datasets

language models scaled with web-scale text. 

vision models got ImageNet. 

now Voxel51 and NVIDIA are trying to do the same thing for physical AI 

autonomous vehicles, humanoid robots, anything that moves around atoms.

building these systems means millions of hours of testing in high-fidelity sims before you can actually deploy anything. 

petabytes of multimodal sensor data from various modalities

 LiDAR, cameras, radar, IMUs. 

but over 50% of simulations end up completely unusable because the input data is bad.

even small calibration or timing errors between sensors cascade through your neural reconstructions. 

you waste weeks and burn through millions in compute costs before you realize what went wrong.

the new FiftyOne Physical AI Workbench. it integrates with NVIDIA Omniverse and NVIDIA Cosmos and basically sits at the start of your sim pipeline to validate, clean up, and prep your multimodal data before it hits neural reconstruction and synthetic data generation.

what it does:

- automatically audits physical AI inputs across 75+ critical checkpoints

- transforms raw sensor streams into structured, searchable datasets

- triggers accurate neural reconstructions and generates synthetic scene variations

this means fewer wasted compute hours. 

iteration loops go from weeks to hours. 

you build safer, more reliable systems on data you can actually trust.

read more about what we've been building: https://lnkd.in/gC6_zn-X

i recently integrated 4 ocr models into fiftyone as remote zoo models these handle text extraction and document parsing. 

all available as remote zoo sources, you can get started with a few lines of code

different approaches for different needs:

1. mineru-2.5

1.2b params, two-stage strategy: global layout on downsampled image, then fine-grained recognition on native-resolution crops. 

handles headers, footers, lists, code blocks. strong on complex math formulas (mixed chinese-english) and tables (rotated, borderless, partial-border).

good for: documents with complex layouts and mathematical content

https://lnkd.in/dSwimdvp

deepseek-ocr

dual-encoder (sam + clip) for "contextual optical compression." 

outputs structured markdown with bounding boxes. has five resolution modes (tiny/small/base/large/gundam). gundam mode is the default - uses multi-view processing (1024×1024 global + 640×640 patches for details). 

supports custom prompts for specific extraction tasks.

good for: complex pdfs and multi-column layouts where you need structured output

https://lnkd.in/dhJn2h-v

olmocr-2

built on qwen2.5-vl, 7b params. outputs markdown with yaml front matter containing metadata (language, rotation, table/diagram detection). 

converts equations to latex, tables to html. labels figures with markdown syntax. reads documents like a human would.

good for: academic papers and technical documents with equations and structured data

https://lnkd.in/dNdTQYKg

kosmos-2.5

microsoft's 1.37b param multimodal model. two modes: ocr (text with bounding boxes) or markdown generation. automatically optimizes hardware usage (bfloat16 for ampere+, float16 for older gpus, float32 for cpu). handles diverse document types including handwritten text.

good for: general-purpose ocr when you need either coordinates or clean markdown

https://lnkd.in/dCTfGfV8


two modes typical across these models: detection (bounding boxes) and extraction (text output)

i also built/revamped the caption viewer plugin for better text visualization in the app: 

https://lnkd.in/dBebWdga

i've also got two events poppin off for document visual ai:

- nov 6 (tomorrow) with a stellar line up of speakers: https://lnkd.in/gkBEY4_r

- a deep dive into document visual ai with just me: 

https://lnkd.in/dFXPimVc

Open article: Document Visual AI with FiftyOne - November 14, 2025 by voxel51.com

Document Visual AI with FiftyOne - November 14, 2025

hype for this new integration, and proud of the team for shipping it at lightning speed. big ups to Daniel and AdonAI for goin beast mode on this one

🚀 Introducing Physical AI Workbench — integrated with NVIDIA Omniverse NuRec and NVIDIA Cosmos to solve the biggest bottleneck in physical AI.  

More than half of all physical AI simulations fail — not from bad models, but from bad input data. A timestamp off by milliseconds or a calibration drift of a few degrees can cascade through your neural reconstructions, leaving you with unusable results after weeks and millions in wasted compute. 

FiftyOne Physical AI Workbench sits at the start of your simulation pipeline and provides turnkey access to NVIDIA technology to ensure every simulation run begins with trustworthy data: 

✅ Audits 75+ critical checkpoints to ensure every reconstruction starts with validated inputs 
🎯 Enriches raw sensor streams into structured, searchable datasets 
🔗 Integrates directly with NVIDIA Omniverse and NVIDIA Cosmos for high-fidelity reconstruction and synthetic data generation. 

The result: an end-to-end workflow that prevents downstream failures — cutting iteration cycles from weeks to hours and eliminating GPU waste on failed reconstructions. We’re excited to see this workflow power autonomous vehicles NVIDIA DRIVE and robotics NVIDIA Robotics development teams move faster with higher accuracy. 

🔗 Read the full technical breakdown: https://lnkd.in/gTy5-TNF

NVIDIA

Product Research Engineer, Deep Learning

AI Makerspace

GitHub Org: https://github.com/AI-Maker-Space

Co-founder & CTO

FourthBrain

Instructor and Curriculum Developer

Founding Machine Learning Engineer

Explorer Hop

Data Scientist

Turtle.AI

Worked on a team of 3 to build an end-to-end ZSL (zero-shot-learning) model for setting and detecting custom wake words!

"Turtle.AI" Wake Word Project

I am obsessed with Machine Learning, Generative AI, LLMs, and more — and I love sharing knowledge as much as I love learning!

Chris "The Wiz 🪄" Alexiuk

Training a model family usually involves a massive trade-off: do you burn the compute to train Small, Medium, and Large variants separately, or do you settle for just one size?

The chefs at the NVIDIA Research kitchen just cooked up a fascinating solution to this bottleneck.

It’s called Nemotron Elastic, and it introduces a "Many-in-One" framework for training reasoning LLMs. Instead of training three separate models, the team trained a single 12B parent model that contains fully functional 9B and 6B submodels nested inside it.

The efficiency stats on this are pretty dope:

- 360x cost reduction compared to training a model family from scratch.  
- 7x cheaper than current SoTA compression techniques.  
- Zero-Shot Slicing: You can extract the smaller models instantly for deployment without any additional fine-tuning.
  
What’s really cool here is that they applied this to hybrid Mamba-Attention architectures. They solved the challenge of keeping the long-context capabilities required for reasoning tasks, even while compressing the model.  

Huge congrats to the research team for another banger, Ali Taghibakhshi and team killin' it as usual.

📄 Read the full paper here: ​https://lnkd.in/gGB8-_kW

AI @ NVIDIA - 🏗️ Build | 🚢 Ship | 🚀 Share

Just a reminder that NVIDIA has an open model for just about any task you could imagine.

Not just models, either. Datasets, code, technical reports, and more.

I've never been part of an organization that ships as much as the Team Green researchers and engineers.

https://lnkd.in/g2xbMXmB

Open article: nvidia (NVIDIA) by huggingface.co

"How do I get hired at XYZ org?"

There's 3 things I would prioritize above all else if you're looking to work in the AI space:

1. Work on your communication, no matter how good you think you are.

I learn something knew when it comes to communicating every single day. Either how to communicate with a new collaborator, or how to communicate my ideas better, or how to communicate more clearly, you get the point.

2. Please, for the love of God, learn your "handshakes".

Code interview? That's a handshake - if you know the handshake, you're more likely to get it. Random questions you feel are unrelated to the job? Another handshake, they show you can learn what you're expected to learn quickly, and thoroughly, which is what you'll need to do on the job. Behavioural interview? Believe it or not, that's a handshake. Don't put on an act, be who you'll be day in and day out, there's a company that is aligned with your work-style and vibe.

3. Understand the tech stack, and know how it fits into the ecosystem.

If you're applying to say, an AI software role at NVIDIA, and you don't know what Nemo is - it's just so much harder to explain and convince the hiring manager how you fit in, and how you're uniquely equipped to help. If you know the stack, and are passionate about some piece of it, you have a way easier time convincing someone that you belong doing that thing.

It's tough out there, with a load of talent, and AI tools complicating everything - but I truly believe that these things can help you stand out, and help you demonstrate why you're the right fit for whatever role you're after.

I know these things worked for me, and I hope they work for you as well. Stay sane out there!

Motivation for this post: I've gotten an influx of people asking me for a referral to NVIDIA recently, and while I don't refer people unless I'm familiar with their work - I do want to ensure that I give something of value to people who want to work at a tech company, and aren't sure where to get started or how to get noticed.

Google DeepMind  released Gemini 3 today, and I gotta say - I feel about this model the same way I feel about every Gemini model: "mid" (except on multimodal tasks, where it really shines bright - especially this time around.

The benchmarks are great, the hype train has definitely left the station and if chuggin' along full steam ahead - but to actually *use* the model leaves me a bit...whelmed.

I've tested it on a number of tasks and apps that I like to run new models through, and honestly it's just kind of another frontier model.

Also, it made this cute SVG of an otter riding a surfboard - and officially made Simon Willison bump the difficulty of his "pelican on a bicycle" test - so, it can't be that bad!

Happy Claude-mas to all that celebrate! Anthropic has done it again, and released yet another model that feels great to use - and crushes code tasks.

Opus 4.5 shows that we've nearly saturated SWE-bench verified (including in multilingual coding), and it's safer than ever. 

They've also added the "reasoning effort" feature we're seeing everywhere else (did Nemotron start the trend? No, but it was one of the first!) - letting you choose (through the API) if you want big think or small think!

Additionally, they did some updates to their apps - classic stuff. 

Overall - this model *feels* great to use, it is great at coding tasks, and it's not too slow for the performance. 

I'm happy to finally have a new daily driver model.

Today was a big day! Ai2 released the newest OLMo model - which is absolutely openmaxxed. OLMo 3, and the accompanying.... everything, is well worth digging through. There are a lot of great tips for all stages of training, data curation, how to think about architecture, and more.

As well, Google DeepMind dropped Nano Banana Pro, which is basically just "in every way better" Nano Banana.

It's been a busy week, and the space shows no signs of slowing down.

This is awesome!

If you want a similar, but extended version of this with a tutorial - check out this blog we wrote:

https://lnkd.in/gdhPJhAx

It's everything you need to know to get cookin' with gas!

AI Developer Experience at Google DeepMind 🔵 prev: Tech Lead at Hugging Face, AWS ML Hero 🤗 Sharing my own views and AI News

Here is the full Python code to build an simple CLI AI Agent from scratch using Gemini 3 Pro.

Build your own: https://lnkd.in/ewKj9uzp

taplio

The Battle

Harpreet Sahota 🥑

Chris "The Wiz 🪄" Alexiuk

Followers

Posts

Engagements

Likes

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance

Performance ScoreChris "The Wiz 🪄" Alexiuk's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Harpreet Sahota 🥑

Chris "The Wiz 🪄" Alexiuk

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Harpreet Sahota 🥑's LinkedIn Performance

Performance ScoreChris "The Wiz 🪄" Alexiuk's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance