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

MongoDB

Staff Developer Advocate (AI/ML)

Slalom Build

Machine Learning Architect

NVIDIA

As a Contributing Writer for NVIDIA, I write and creatte compelling content on Machine Learning, Data Science and AI topics for a data science audience. 

My role involves researching and staying up-to-date on the latest advancements in these fields and identifying and writing about topics that interest our readers.

Key Responsibilities:

- Researching and staying up-to-date on the latest advancements in Machine Learning, Data Science and AI
- Identifying and writing about topics that would be of interest to our readers
- Creating compelling and informative content for our data science audience
- Collaborating with other writers and editors to ensure that our content is of the highest quality
- Leveraging my expertise in these fields to create technical tutorials and explain complex concepts in an easy-to-understand manner

Contributing Writer

Emeritus

Providing educational services for a Professional Certificate in Machine learning and Artificial Intelligence for Imperial College Business School

- Creating and presenting lectures on Data Science and Statistics topics in relation to machine learning
- Informing students on the latest advanced trends within the field of deep learning and artificial intelligence
- Creating and presenting lecture content around machine learning and deep learning topics such as convolutional neural networks, backpropagation, neural networks etc
- Introducing and showcasing functionalities of tools utilised within the ML field to students, such as Keras, TensorFlow, SciKit Learn, Pandas etc
- Educating students on the development of machine learning models, from traditional Random Forest classifiers to more complex deep learning neural networks architecture.
- Informing students about real-world problems where machine learning solutions can be leveraged
- Marking and providing grades and feedback on weekly assignments and a final course capstone project.
- Weekly delivery of machine learning and deep learning course materials over 25 weeks.

Program Lead / Learning Facilitator 

Loveshark

Leading Computer Vision initiative and implementation for unique and immersive experience within iOS mobile game applications.

Responsible for the research and implementation of common Computer Vision techniques such as pose estimation, semantic segmentation, gesture recognition and face/facial feature detection.

Mainly working with the TensorFlow machine learning library (Python and JS) and Apple's CoreML models

Computer Vision Engineer

O'Reilly Media

Constructed, led and delivered a training course on Computer Vision in the O'Reilly Live training session.

https://bit.ly/3JoOjAi



- Led a team of students in learning the theory and practical components of computer vision and deep - 
learning
- Provided explanations of widely used terminologies and taught Python programming language
- Guided students in implementing techniques and solutions to a typical computer vision image - classification problem using Python, TensorFlow, NumPy, and pandas
- Used Jupyter notebooks for hands-on learning and practice
- Taught the application of computer vision in practical environments
- Explained the fundamental components of deep learning as a machine learning field
- Demonstrated how to leverage deep learning to solve computer vision tasks
- Taught the structural components of deep neural networks and convolutional neural networks
- Provided guidance for students to implement solutions to common computer vision tasks
- Demonstrated the use of machine learning libraries to implement deep learning solutions
- Helped students to build a deep neural network that classifies images
- Helped students to build a convolutional neural network (AlexNet) that classifies images
- Demonstrated the use of TensorBoard to visualize the neural network training process

Course Instructor

Eventcide

Founder and Developer

Allianz

Front End developer utilizing the Angular 1.x and Angular framework

Contract Web Developer

Front End Developer focused on the Angular Framework and JavaScript Web applications

TeacherHQ

Developing and maintaining the application Frontend with AngularJS
Developing and maintaining the application Backend with NodeJS and Express
Developing and maintaining the application database (PostgreSQL)(KnexJS)
Deploying and maintaining application and database server on AWS
Developing functional testing and maintaining Front End browser testing with Nightwatch
Implementing and handling application error logging with Rollbar
Building RESTful APIs endpoints
Managing ongoing project plans and timelines.
Training junior developers.
Turning ideas from the team into product features
Oversight of core platform development (Front and Back end)

Full Stack JavaScript Developer

Richmond Alake is a highly experienced Machine Learning Architect and Engineer with over five years of expertise in the field. He specializes in Computer Vision and Deep Learning and has a proven track record of successfully developing and integrating deep learning models to solve a wide range of problems, such as motion detection, object detection, and pose estimation. Throughout his career, he has worked with a diverse range of clients, including large conglomerates, financial institutions, and small startups. In addition to his professional work, Richmond also serves as an AI advisor to a number of startups in the UK and the US.

With a background in building websites and mobile applications, Richmond is a firm believer in using technology to solve everyday problems. He has extensive knowledge of Machine Learning and has written over 200 articles on the subject, gaining over a million views. He was recognized as one of Medium's top AI writers in 2020/2021 and has collaborated with companies such as O'Reilly, BuiltIn and Nvidia to develop effective educational and informative learning materials on AI.

Currently, Richmond Alake is a Machine Learning Architect at Slalom Build UK. As the first hire of the machine learning practice in the UK division, he is responsible for helping organizations move from machine learning research to productionization and assisting maturing organizations in promoting AI models into existing infrastructure to drive commercial and business value. His main role as an ML Architect is to assist organizations in developing and maintaining machine learning pipelines by implementing MLOps principles, techniques, and tooling. He is well-versed in Feature Stores and has conducted internal training for Data Engineers, Data Scientists, and ML Engineers.

Richmond Alake

Is it safe to say I coined the term "Memory Engineering"
At least within the context of AI/Agentic systems.
Someone let me know who beat me to it 😉

Either way I'm glad to see Agent Memory taking centre stage in a number of conversations in Agent Engineering. And equally glad to see some startups raising funds to solve what argualbly is one of the last problems to solve in AI.

Want to hear more about Memory Engineering,
Then you can catch my talk at the AI Native Dev Conference in NY next week.

And an advance warning: Nothing in my talk will be “new”.
Vectors are not new.
Automated workflows are not new.
And BTW AI existed long before ChatGPT.

Those who get to listen in on this talk will understand the evolution of the job AI Developers are doing to make AI reliable, believable and capable. 
I'll touch on the following:
1️⃣ Prompt Engineering as Maximisation (pack more tokens into the context window)
2️⃣ Context Engineering as Optimisation (select and structure the right tokens)
3️⃣ Memory Engineering as Continuation (persist, retrieve, adapt, and improve over time)
4️⃣ My developing understanding of Continual Learning
5️⃣ The work AI Developers building AI Agents will be doing the most in 2026 

If you’re in New York next week, say hi.
If you’re not — I’ll share the key insights here as always.

The AI Stack Engineer | AI Strategist and Educator

I must be one of the worst "LinkedIn Influencers"
I didn't even notice I passed 10,000 followers.
Apparently it's something to be celebrated.
So....let's celebrate....

Now that's out of the way 😉
Follow me if you are interested in hearing about
- AI Agents and Applications development (technical stuff)
- Agent Memory and Memory Engineering (mostly propaganda lol)
- Occasional pics/vids of my travel when I decide to leave my home office
- And summary pancake discussion with my friends and folks in the AI space
- #100daysofAgentMemory and more

When is the next milestone to be celebrated?

How far would you go to prove a point?
For me? 
Pretty far.
Like UK → New York far. ✈️.

🧠 Day 59 / 100 of Agent Memory

Today I found myself in front of dozens of AI developers at the AI Native Dev Conference in NYC(a moment that’s been over a year in the making).

I’ve spent the last couple of years advocating for a clear, practical way for AI developers to implement Agent Memory — something that’s intuitive, production-ready, and doesn’t require a PhD in retrieval pipelines.

Not a lot of people understood what I meant when I said the words "Agent Memory" or "Memoru Engineering"... it's obvious now... back then, not so much.

So I built MemoRizz. 
1/3 serious mission. 1/3 wild experiment. 
1/3 running joke that somehow escaped.

But very quickly — talking to real engineers working on real production agent workloads — I realised something: Their biggest bottlenecks weren’t model hallucinations. It was retrieval, latency, chunking, indexing, search optimization, and hybrid retrieval. All the “deep plumbing” work that makes agents reliable.

So today, I delivered a 2-hour workshop showing developers not just what Agent Memory is, but how to build it using a fully integrated Agent Stack:
🔹Oracle AI Database — Memory Provider (Memory Core)
🔹Tavily — Web search/internet access
🔹OpenAI — LLM (Reasoning Core)/Embedding Model provider

Code is here: https://lnkd.in/e9WEZUdZ.

We even took a crack at implementing Deep Research: https://lnkd.in/e_Yhv-ZJ

And after the workshop, one developer came up to me and said:
“Wow Richmond… I didn’t know you were going into that much detail.”
Look — same. That makes two of us.

But the more I teach this stuff, the clearer the whole Agent Stack becomes:
And right in the middle — quietly powering all the continuity, personalization, and intelligence — is the Memory Core.

More on Memory Core later.

Side note:
I love talking about Agent Memory…
But man, I think I’ve been to the US like five times this year 😅
What conferences should I be speaking at in Singapore and Tokyo next? I think it’s time I spread this Agent Memory gospel to the other side of the world.

Special thanks to the Tessl team for making this happen and Sofia Guzowski for the pics.

#100DaysOfAgentMemory #AgentMemory #MemoryEngineering #AIEngineering #AgenticSystems #OracleAIDatabase

Google is onto something here.
And yes. This is more Agent Memory propaganda.
Buckle up 🙂

🧠 Day 58 / 100 of Agent Memory

If you are interested in agentic systems, context management, memory architectures, multi-agent coordination, or the future of AI engineering, then bookmark this:

“Context Engineering: Sessions, Memory” [1]
 Authors: Kimberly Milam, Antonio Gulli, and a full research team.

This is a 70-page deep dive into: Sessions. Memory managers, Retrieval strategy, Compaction, Multi-agent interoperability, Procedural memory, Memory provenance, Memory-as-a-tool, and Production architectures

That's a lot.

But if you've been following the #100daysOfAgentMemory series, then you will be able to follow what most of the whitepaper introduces and explains.

I’m not going to break it down here (that deserves its own day).

But I'm glad, like I said
2026: The year of AI/Agent Memory
How many Agent Memory Talks can I do in a year?

Whitepaper: https://lnkd.in/eGP6amkq

#100DaysOfAgentMemory

Are the Cursor team secretly following the Agent Memory series? 👀

🧠 Day 57/100 of Agent Memory

Because back on Day 53[1] of the Agent Memory series, I speculated that Cursor’s internal architecture probably used a hybrid retrieval approach — blending pre-computed (RAG-style) retrieval with Just-in-Time (JIT) context loading to make their coding agent more efficient.

And last week, they published “Improving Agents with Semantic Search”[2] — confirming almost exactly that. They revealed how semantic search, paired with traditional grep-style keyword search, dramatically boosts coding agent performance:

✅ Up to 12.5% higher accuracy
✅ More retained code changes
✅ Fewer iterations to reach a correct solution

💡 The Not-so Hidden Shift: From Search to Continual Learning
It looks like Cursor’s team didn’t just optimize retrieval — they turned it into a learning process. By training their custom embedding model on agent session traces, they’ve essentially built a retrieval feedback loop.

Each interaction becomes an example of what the agent should have retrieved earlier, helping the system refine its sense of semantic relevance over time.

That’s a subtle but major step forward — retrieval as experience accumulation.

In traditional systems, retrieval is a stateless function: you query, you fetch, you forget. But in Cursor’s approach, retrieval now evolves — using history to inform future behavior. That’s the foundation of memory engineering.

And what’s especially important here:  even those “small” percentage gains (like 0.3%, 1%, or 2%) aren’t trivial.  In large, production-scale engineering environments, they compound into huge outcomes — fewer bugs, higher code reliability, smoother merges, and stronger developer trust.

🧠 Rethinking Continual Learning

For the longest time, I viewed continual learning as something that applied only to large language models — complex, expensive, and risky to run in production. But this approach opened my eyes: embedding models might be the low-hanging fruit for continual learning. 

They can learn from retrieval logs, feedback loops, or developer traces — improving contextual accuracy without retraining the full LLM.

They’re smaller, faster, and much more adaptable.

That said, there’s a cost curve:
1️⃣ Reindexing overhead rises as embeddings drift, demanding more frequent vector rebuilds.
2️⃣ Forgetting risk increases if updates aren’t properly regularized.

So while embedding-level continual learning gives agility, it has to happen at a measured cadence — perhaps monthly or quarterly, balancing freshness and stability. 

[1] https://lnkd.in/esxqrGWp
[2] https://lnkd.in/eDniUfHT
[3] OpenSpeech Notes: https://lnkd.in/eT4mAVKA

#100DaysOfAgentMemory #MemoryEngineering #ContextEngineering #ContinualLearning

⚡️ Day 56/100 of Agent Memory 🧠
 Speed Isn’t About Shipping Fast 
— It’s About Delivering Value When It’s Needed.

One of the biggest advantages in AI today is speed.
And I don’t mean the speed of building features, or even the speed of raising millions. I mean the speed of delivering value — exactly when it’s needed — to your customers and your community.

Let me tell you what happened 10 hours before I had to go live and teach a five-hour session to nearly 900 AI developers on AI memory management and evaluation.

I was running my demo code — it used Galileo(code in at end of post) to show AI evaluation pipelines in action. Except… it wasn’t working. Nothing breaks quite like code right before a live event 😅

I immediately pinged Conor Bronsdon at Galileo. It was a long shot. He’s busy — talking to partners, teams, customers.

 Still, I sent the message:
“Hey Connor, this is a long shot, but I’ve got a five-hour course tomorrow with 900+ developers. I want to show them how to evaluate AI agents and RAG systems using Galileo. I’ve used Galileo before — I know the value — and I really want to show it properly.”

Within 30 minutes, Connor replies: “Richmond, how can I help? Give me your number — I’ll call you.” Minutes later, I’m connected to Jim Bennett. We jumped on a call that same evening. 

Jim spent sometime debugging, testing, and running my code on his end until everything worked flawlessly. The issue had to do with my development environment and the retention of environment variables that were meant to not there.

🔟 hours later, I delivered a 5️⃣-hour course showing developers how to build reliable, believable, and capable AI systems — powered by Galileo’s evaluation, logging, and tracing capabilities.

That, to me, is speed.
 Not a faster sprint cycle. Not another “launch in public.”
 Speed is how fast you can help others create value.

🧠 Why Evaluation Matters — Especially for Memory Systems
When we talk about Agent Memory, we’re really talking about systems that must learn, recall, and act consistently over time.  And that means they must be evaluated — continuously. AI evaluation is the discipline that ensures your system doesn’t just work once, but works reliably over time.

It spans multiple layers:
Information Retrieval Evaluation → Are you retrieving the right chunks?
RAG Evaluation → Are the retrieved chunks useful and do you faithfully grounded responses?
Agent Evaluation → Does the agent act consistently, rationally, and contextually across tasks? Does it select the right tools?
Memory Evaluation → Does the agent recall, adapt, and reflect with reliability over multiple sessions?

Massive shout-out to Connor and Jim at Galileo for proving that the best teams don’t just build tools —They help others build better systems.

Notebook: https://lnkd.in/ebHJTjid

#100DaysOfAgentMemory #AIEvaluation #MemoryEngineering  #AgenticSystems #RAG #LLMs

taplio

The Battle

Harpreet Sahota 🥑

Richmond Alake

Followers

Posts

Engagements

Likes

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance

Performance ScoreRichmond Alake's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Harpreet Sahota 🥑

Richmond Alake

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 ScoreRichmond Alake's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance