Armand Ruiz

Director Data Science & AI Elite

VP of Product - AI Platform @IBM

As the Director of AI Engineering at IBM, I lead the watsonx team, a global group of over 1,000 AI Engineers who work on cutting-edge Generative AI solutions for various industries. With more than 12 years of experience in AI, I have a passion for democratizing practical AI skills and empowering professionals to leverage AI to transform their business.

Previously, I managed IBM's Data Science and AI Elite, delivering complex AI implementations for hundreds of clients. 

I also led the product management of IBM Watson Studio and Watson Machine Learning, an AI platform used by hundreds of customers and SaaS apps powered by AI. I hold a Master of Science in Telecommunications from Universitat Politècnica de Catalunya and speak three languages: Spanish, Catalan, French, and English.

Connect with me!

✉️ JOIN MY NEWSLETTER: Every Saturday morning, you'll get 1 actionable tip to launch and grow your AI strategies in business in less than 4 minutes. → GO HERE: https://www.nocode.ai/

📧 Email: armand@nocode.ai
X: @armand_ruiz


"It is kind of fun to do the impossible"
All opinions are my own and don't represent my employer.

Director watsonx team - Client Engineering

I am building the world's largest team of 𝐀𝐈 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐬 by

- Recruiting hundreds of new engineers around the globe
- Reskilling our existing talent
- Offering them top-quality training in new-generation AI

Responsible for Pre-Sales Generative AI Pilots with the new IBM watsonx.ai Platform.

I’ve been practically living in Claude Code lately. If you build with AI daily, you know that your orchestration workflow is just as important as the model itself.

Recently, Boris Cherny (the creator of Claude Code at Anthropic) shared the internal best practices and workflows his team actually uses on a daily basis. 

Someone brilliantly distilled those threads into a structured CLAUDE.md file that you can drop straight into the root of any project.

It essentially acts as a system prompt that turns Claude into a much more autonomous, rigorous engineering partner.

Here is what the file enforces:
1/ Workflow Orchestration: Mandates a "Plan Node Default" for any task over 3 steps and uses subagents liberally to keep the main context window clean.

2/ The Self-Improvement Loop: This is the real magic. After ANY correction, it updates a tasks/lessons.md file. You are building a compounding system where the mistake rate drops over time because it actively learns from your feedback.

3/ Verification Before Done: It cannot mark a task complete without proving it works (diffing behavior, running tests, checking logs). "Would a staff engineer approve this?"

4/ Autonomous Bug Fixing: Zero hand-holding. You point it at failing CI tests or error logs, and it goes to work without requiring constant context switching from you.

5/ Strict Task Management: Forces a "Plan First" approach written to a todo.md with checkable items before implementation even starts.

It forces the AI to prioritize simplicity, find root causes instead of temporary fixes, and minimize the blast radius of its changes.

If you are spending hours a day in the terminal with AI, setting up a strong .md instruction file like this will save you a massive amount of time and frustration.

It takes time and effort to set this up. If you do, you are ahead of everyone else.

If your employer isn’t providing unlimited Claude Code + secured MCPs to connect your enterprise repos, email, work-chat, meeting AI transcripts, calendar, and docs... they are likely moving at legacy speed. 

Without that context and tool access, you’re stuck in a cycle of manual copy-pasting and reinventing the wheel while the rest of the market moves in parallel workflows.

My advice? Don’t wait for the corporate roadmap to catch up. If you aren't getting these tools at work, force yourself to learn these agentic workflows on the side. 

When companies move this slowly, they eventually run into trouble and you don't want to find yourself at a disadvantage in today's market because your professional environment kept you in the past.

Stay curious and keep your skills ahead of your company's tech stack.

Everyone's talking about which LLM is best for agents. But after reading Lance Martin's deep dive into Manus's architecture, I'm convinced the real differentiator isn't the model. It's the context engineering around it.

The problem: A typical Manus task makes ~50 tool calls. Without careful management, results pile up in the context window and the LLM gets dumber the longer it runs. 

Anthropic calls it "attention budget depletion." 

Chroma calls it "context rot."

Manus solves this with three strategies:
𝟏. 𝐑𝐞𝐝𝐮𝐜𝐞: Older tool results get compacted to file path references. When that's not enough, the full trajectory gets summarized via structured schema.

𝟐. 𝐎𝐟𝐟𝐥𝐨𝐚𝐝: Fewer than 20 atomic tools bound to the LLM. Everything else runs in a sandbox via bash. The model stays lean.

𝟑. 𝐈𝐬𝐨𝐥𝐚𝐭𝐞: Sub-agents exist to give each task a clean context window, not to mimic human roles. No context bleed.

On top of that, Manus doesn't commit to one model. It routes tasks; Claude for code, Gemini for multimodal, OpenAI for math. 

Model-agnostic by design.

The takeaway: Manus has been refactored 5 times since launch. If a stronger model doesn't improve your agent, your engineering is the bottleneck.

Stop obsessing over which model to pick. Start obsessing over the implementation around it.

Full article: https://lnkd.in/gaY3rAV5

Prediction: Token budgets will soon surpass human salaries.

We are entering an era where the cost of compute for reasoning is becoming a primary line item, and the numbers are scaling fast.

I am already seeing developers consuming $100k+ a week in tokens.

It sounds astronomical, but when you look at the output, the ROI is there. They aren't just running a single chat thread; they are spinning up massive parallel processes. Swarms of agents coding, debugging, testing, and architecting simultaneously. The velocity is incredible.

But at that scale, cloud inference costs become the bottleneck.

The big unlock? Local Compute.

Putting a Mac Mini (or equivalent high-performance local hardware) on every developer's desk to run local models is the inevitable next step. We need to move the heavy lifting to the edge to make this sustainable. 

Invest $10,000 on a local computer and you have unlimited AI.

This creates a massive new need for the Enterprise: Fleet Management for Models.

Just as we manage laptops with MDM, companies will need a way to manage these distributed AI systems; securely pushing the latest Open Source models to local machines, updating weights, and orchestrating the workflow between local "thinking" and cloud "verification."

The future is hybrid, and the infrastructure to support it is being built right now.

I’m building my Personal OS to automate my mental overhead. For now, it’s simple, but the potential is already showing. I’ve moved away from "checking notifications" to a Pulse-based system.

How it works: 
1️⃣ Orchestration: Claude.md acts as the central hub, invoked by a launchd scheduler. 
2️⃣ Sequential Logic: Every "Pulse" (Morning/Midday/EOD) triggers a sequence of my agents: Scout → Horizon → Focus. 
3️⃣ Multi-Channel Delivery: I get a unified briefing via macOS notifications, an HTML digest, and my terminal.

The Agent Roster:
- Scout (/scout): Google Chat & Email monitor. It knows my VIP list and drafts replies before I even see the message. If something is important it will notify me, and even give me a call.
- Horizon (/horizon): The calendar expert. No more walking into meetings cold.
- Focus (/focus): The filter. It turns the noise into a top actionable list.

I feel spend less time "managing" my day and more time executing it. 

I also turn all the context for the day ahead into a fun podcast, and that gets me ready for the day and fully prepared.

I've been using computers in a completely different way lately. No much clicking. No browsing through tabs. No filling out forms. I describe what I want done, and agents go do it.

Many are still building software for humans. Replicating existing apps with AI. 

But that's not where this is going. The moment I stopped thinking of myself as a "user" and started thinking of myself as an orchestrator, everything changed. I set goals. I define constraints. I verify outcomes. That's it. The entire middle — the actual work — I don't touch it anymore.

And now I can't unsee it: most software we're building today is already obsolete. It's designed for humans to operate. But humans aren't going to be the operators anymore.

The new generation of software won't have dashboards. It won't have onboarding flows. It'll have APIs, state machines, and feedback loops built for AI agents, not for people.

The real question isn't "how do we develop this product/feature for humans using AI?" --> It's "if no human ever had to touch this, what would we build from scratch?"

if you are in a leadership position, you are punishing your team by keeping them busy.

If your team is too busy doing their "normal jobs" to spend hours a week experimenting with AI, you are failing them.

You are setting them up to be replaced by people who have the time to learn.

They will get disrupted, either by another department within the company or by a competitor, not because of AI, but because of your lack of urgency and understanding of the current landscape.

You have to give people the time, because integrating AI into daily workflows isn't something that happens overnight.

You have to actively remove work from their plates so they can learn the new way to work.

Because we are now all beginners again.

If you don't give them the time, they will never learn it, and your company will suffer a slow, painful decline into inefficiency.

No matter how big you are.

you need to clear your calendar and invest in AI implementation for your own productivity asap

The hardest part isn't the technology. It's making the time to actually sit down and try it.

Block 2 hours a day. Thank me later

Parallelism is the biggest unlock in AI. The real shift isn't speed; it's concurrency.

 I can now run 10 "hard things" simultaneously while I'm in meetings.

My workflow has flipped:
- Old way: Block time to do the work. Maybe use AI to do it faster.
- New way: Block time to review the results and give more context knowledge to the AIs.

We are moving from being "doers" to orchestrators. Your output is no longer limited by your hours, but by the number of parallel streams you can manage.

I get 100s of emails and chats daily. The only way to manage it is to stop everything, dive in, and manually filter the signal from the noise. It's inefficient and a total focus-killer.

So i built Mark, my Communications AI Agent. Its job is simple:
1/ Monitor Everything: It watches all my inboxes and chats 24/7.
2/ Surface What Matters: It identifies critical messages and ignores the rest.
3/ Notify Intelligently: I get a clean Terminal and Desktop notification with a summary file. No more context switching.

But that's just the start. The real magic is next → I'm connecting Mark to a shared knowledge base of all my documents, past conversations, and daily meeting notes.

Soon, it won't just notify me. It will draft context-aware replies for almost everything.

That's my way to solve this. I'm curious...have you developed something similar?

I moved from the EU to the US more than 10 years ago, and while I miss many things about Europe, I have to agree with the man of the hour 🦄 Peter Steinberger : the US is an incredible place for builders right now. 

There's an undeniable energy and enthusiasm here that's hard to find elsewhere. 

It feels like a special time to be building here.

Manus just shared why most people are still stuck in the "chat" era: the real value has shifted from conversational bots to custom superagents.

Chatbots answer questions; superagents execute workflows. 

But they aren't magic, you need to understand the underlying mechanics and put in the time to customize them for your specific needs. If you treat them like a standard search bar, you’re leaving 90% of the value on the table.

Here is how to actually get them to work for you:
1/ Move from Prompting to Engineering: Stop asking "Can you do X?" and start defining the architecture of the task.

2/ Invest in the Setup: Spend the time upfront to define personas, data sources, and specific tool-use constraints. The more "narrow" the agent's focus, the higher the quality.

3/ Iterative Refinement: A superagent is never "done." You have to monitor the execution and tune the instructions based on real-world failures.

Swipe through for the tips that actually matter for high-stakes execution.

Manus Intro

What’s the point of buying software when you can just create it on the fly?

It might take our generation a minute to adapt, but the next generation is growing up expecting to build exactly what they need. For them, the only thing that matters is the infrastructure to power it.

I’ve been building non-stop lately. I’ve never felt this empowered in my career. I can literally turn my ideas into reality in a matter of hours.

yesterday at the gym I heard someone pitching their new AI tool that helps you “do more, faster.” i feel like doing less and slower. am i the only one?

ok, so you moved from a chatbot to an AI agent. and you think you are ahead already.

but if you are still the one triggering all the actions...

if the agent isn't working autonomously for you all day long...

and if you don't have 10 of them running in parallel...

you are nowhere close to calling yourself AI-native. 

i recommend you spend the weekend learning. this is a good place to start: https://lnkd.in/eW4Sam_6

Every PM should understand evals. Not "be an expert." Not "build the eval framework yourself." Just understand them.

Here's why.

A/B testing changed product management forever. Before it, PMs guessed. After it, they measured. Evals are doing the same thing for AI products and systems but most PMs have heard the word but they don't know what they're talking about. 

An eval is how you answer one simple question: "Is my AI actually good?"

Not vibes. Not "it seems better." Not "the demo looked great." Actual, measurable, repeatable quality assessment.

Here's what you should tell tell your product leads:

1. Start with a golden set: a curated dataset where you know the right answers. This becomes your source of truth.

2. Define your graders: what does "good" look like? Accuracy? Tone? Factual correctness? Be specific.

3. Run your model against the golden set and score it. That's your baseline.

4. Every time you change something - new prompt, new model, new data — run evals again. Did it get better or worse?

5. Automate this. The speed of your eval loop determines how fast your AI improves.

That's it. That's the core.

The teams that ship reliable AI products aren't the ones with the best models. They're the ones with the best evals practice.

If you're a PM working on anything AI-related and you can't answer "how do we know it's good?"... start there.

Evals are the new A/B testing. And just like A/B testing, the PMs who get it early will have a massive advantage.

Manus Agents just launched on Telegram and it changes how you interact with AI agents entirely. I've been using this all weekend non-stop.

This isn't a chatbot bolted onto a messaging app. It's the full Manus — reasoning, multi-step task execution, tool use — running inside a conversation you already have open.

Here's what makes it different:
1/ Zero setup friction. Scan a QR code. You're connected in under a minute. No API tokens. No config files. No terminal.

2/ Full task execution from chat. Research, data processing, structured reports, PDFs — initiated and delivered without ever leaving the conversation.

3/ Multimodal by default. Send a voice message, a photo, a document. The agent transcribes, interprets, and acts. No manual file handling.

4/ It remembers how you work. Set your communication style — concise, structured, conversational — and it adapts across every interaction.

5/ You pick the model. Manus 1.6 Max for deep reasoning and creative work. Manus 1.6 Lite for fast, everyday tasks. Both go end-to-end.

But here's what got my attention even more — the cofounder just shared what's coming next:

1/ Create your own specialized agents and plug them into any group chat.

2/ WhatsApp, LINE, Slack, Discord support landing soon.

3/ Native Windows and Mac apps that let Manus operate your computer — like Browser Operator, but way more powerful.

And a lot more already in the shipping pipeline. Big updates dropping over the next 30 days.

Your AI agent shouldn't live behind a login screen on a separate platform. It should be wherever you already are — one message away, ready the moment you need it.

Learn more here: https://lnkd.in/gSuAx5UX

trust your AI overlord and... just hit accept 🤣

Recently, I stumbled across a YouTube video that completely stopped me in my tracks.

An urban explorer infiltrates the abandoned IBM center on the Côte d'Azur in La Gaude in South of France. Hearing the narrator describe it as a "vestige of Silicon Valley" and seeing that massive concrete colossus sitting empty in the middle of the forest is absolutely heartbreaking 💔 

For me, that place wasn't just a building. It was pure magic. It was where my professional life truly began and where my life completely transformed.

Early in my career, I had the chance to visit this unbelievable, high-energy campus while I was interning for IBM in Belgium. I knew immediately I belonged there, and I did absolutely everything in my power to get hired and eventually, I earned my spot.

When I look at the footage of those abandoned offices and empty reception halls, I don't see decay. I see endless, vibrant memories:

- Going from a nervous intern to earning my first full-time role.
- Working late into the night and battling those early-career nerves before major presentations.
- Learning a completely new language (French!).
- The incredible lunches, coffee breaks and walks around the campus with colleagues who became amazing, lifelong friends.
- Spending hours on the campus soccer & tennis courts, both competing in tournaments and playing with colleagues. I also learned to play la pétanque!
- And most importantly, it’s where I met my incredible wife.

That campus was the launchpad that gave me the opportunity to eventually relocate to the US. 

Looking back, I can draw a direct line from walking through those doors in France to my time as VP of AI Platform at IBM, and to do now the most exciting work in my career in Silicon Valley.

The doors might be closed, but the energy of what we experienced there is still very much alive. 

I will forever be grateful for that chapter, those people, and those memories.

Link to the video: https://lnkd.in/gA-Wja2s

Open video: INFILTRATION d’IBM, le plus gros site abandonné de la Côte D’Azur mais aussi le plus Surveillé !!!  on youtu.be

INFILTRATION d’IBM, le plus gros site abandonné de la Côte D’Azur mais aussi le plus Surveillé !!! 

I thought I was a heavy AI user. Then my usage 3x'd in the last 30 days. The pace of change is accelerating and what was cutting-edge a month ago is now the baseline. The new default is managing AI agents, not just using them.

This chart below is my AI usage (tokens)

Ranked by Favikon as:
↳ #1 AI Creator on LinkedIn India 
↳ Top 10 AI Creators on LinkedIn Globally 

Featured in Forbes and Wired

For 14 years, I've been obsessed with a single question: how do we teach machines to truly see and understand our world?

This quest began at Qualcomm. My 3D reconstruction technology landed in Forbes, and the object tracker I co-designed was used in millions of mobile devices. When someone casually used my work in their palm, I glimpsed what's possible when research meets reality.

At Intel Labs, I immersed myself in creating AI systems that bridge the gap between theoretical potential and practical performance—systems that could deliver sophisticated intelligence within real-world constraints. Today at Valeo, I lead a team of AI experts developing computer vision for autonomous vehicles. We're teaching cars to go beyond simple object detection toward true scene understanding. Each breakthrough is a step toward fundamentally transforming how people and goods move worldwide. We are reimagining transportation.

Since 2019, I've guided teams at NASA's Frontier Development Lab as their Machine Learning Lead. We've tackled fundamental questions about our Sun and understanding it better. Our breakthroughs have since been published in conferences like NeurIPS. 

Along the way, I saw a problem I couldn't ignore.

I watched brilliant scientists—experts in physics and natural phenomena—blocked from leveraging AI. Not by ability. But by clarity. Incredible AI research remained locked behind jargon. 

There had to be a better way to get the message across. So, Gradient Ascent was born. In it, I break down complex AI concepts for engineers, product leaders, CEOs, and investors making critical decisions shaping our future. Their response has been humbling.

Outside the lab, I'm an award-winning nature photographer and enjoy illustration and music. These pursuits fuel my creative approach to problem-solving.

✍️ Subscribe to my newsletter for insights that transform technical AI jargon into actionable understanding: newsletter.artofsaience.com

———
Note: All opinions are my own and do not reflect the opinions of my employer

Sairam Sundaresan

You don’t need 100 tutorials.
Just these 8 videos on AI agents:

1. LLM Introduction: 
https://lnkd.in/gN8sv7Q5
2. LLMs from Scratch: 
https://lnkd.in/gJt-SQj2
3. Agentic AI Overview (Stanford): 
https://lnkd.in/gk4GKdxa
4. Building and Evaluating Agents: 
https://lnkd.in/ghBiVjdF
5. Building Effective Agents: 
https://lnkd.in/g_m78sid
6. Building Agents with MCP: 
https://lnkd.in/gAzBzr3W
7. Building an Agent from Scratch: 
https://lnkd.in/g9GR9b9F
8. Philo Agents: 
https://lnkd.in/gnxRq9n9

♻️ Repost to help someone master AI
➕ Follow me, Sairam, for AI engineering that works

Want to learn the basics first? Check this out: https://lnkd.in/giV_N6KU

AI Engineering Leader | Author of AI for the Rest of Us | I help engineers land AI roles and companies build valuable products

6 books. 6 courses. 6 months.
Ship real LLMs without the wait or the tuition.

I've spent over 15 years in AI research and engineering.
Read hundreds of papers. Tested countless courses.
Shipped AI products to production.

Here's the truth:
You can get 90% of the practical LLM knowledge you need without spending $120K on a degree.

So, I put together a focused 6-month roadmap for you.
Each month, you'll read 1 book and complete 1 free course.
Just 60 minutes/day.

Here’s your roadmap:

🔸 Month 1: Python Basics

📘 Book: Automate the Boring Stuff with Python
✅ Automate repetitive tasks
✅ Write practical scripts
✅ Use real-world examples fast

🎥 Course: Python for Everybody (Michigan)
✅ Core Python skills
✅ Solve real coding problems
✅ Build confidence through practice

🔸 Month 2: Practical Deep Learning

📘 Book: Deep Learning for Coders (fastai)
✅ Build neural nets quickly
✅ Learn PyTorch workflows
✅ Go from beginner to builder

🎥 Course: FastAI's Practical Deep Learning
✅ Train real models fast
✅ Solve real-world tasks
✅ Minimal code, big results

🔸 Month 3: NLP Mastery

📘 Book: Speech & Language Processing
✅ Core NLP foundations
✅ Deep neural NLP understanding
✅ Learn transformer theory

🎥 Course: Stanford NLP (CS224n)
✅ Master attention mechanisms
✅ Train SOTA models
✅ Work hands-on with BERT

🔸 Month 4: LLMs from Scratch

📘 Book: Build an LLM from Scratch
✅ Understand LLM building blocks
✅ Build your own GPT
✅ Learn training and inference

🎥 Course: Stanford LLM from Scratch (CS 336)
✅ Code LLMs from scratch
✅ Understand evals and scaling
✅ Learn alignment and reasoning

🔸 Month 5: Applied NLP

📘 Book: NLP with Transformers (Hugging Face)
✅ Understand transformer applications
✅ Fine-tune for Q&A, chatbots, more
✅ Apply SOTA NLP tools

🎥 Course: Hugging Face NLP Course
✅ Build transformer pipelines
✅ Understand reasoning models
✅ Deploy real NLP apps

🔸 Month 6: LLMs in Production

📘 Book: LLM Engineer's Handbook (Iuzstin & Labonne)
✅ Design LLM systems end-to-end
✅ Build RAG pipelines reliably
✅ Monitor and maintain in prod

🎥 Course: Stanford MLSys Seminars
✅ Real-world AI infrastructure
✅ Proven deployment patterns
✅ Lessons from top AI teams

That’s your 6-month roadmap to LLM mastery.
All it takes? Just 60 minutes/day of consistent learning.

Tag someone who’s learning AI.
This roadmap will save them months.
Or better yet, take the 6-month challenge together.

🔁 Repost and share this with them.
➕ Follow me, Sairam, to learn AI the right way.

–
Want to go deeper? Gradient Ascent has you covered. Join 24k+ readers from Google, Meta, Netflix, and over 160+ countries worldwide.

P.S: It's ok if you aren't able to finish this in 6 months. The goal is to start and build a habit of deliberate practice and learning. Good luck!

Stanford just made a $200,000 AI education free.
The same lessons that built Silicon Valley are now on YouTube.

📚 Here's your map to go beyond prompting and start building real AI systems:

1️⃣ CS221 – Artificial Intelligence
Foundations of AI: search, reasoning, planning, decision-making
🔗 https://lnkd.in/gM9ndUCf

2️⃣ CS229 – Machine Learning (Andrew Ng)
Supervised/unsupervised learning + deep dive into algorithms
🔗 https://lnkd.in/gDc9siUC

3️⃣ CS230 – Deep Learning (Andrew Ng)
Neural nets, CNNs, RNNs, and best practices for deep models
🔗 https://lnkd.in/guaavVdk

4️⃣ CS231n – Deep Learning for Computer Vision
Neural nets, CNNs, Generative Models, and VLMs
🔗 https://lnkd.in/gVJZ-xMq

5️⃣ CS229M – Machine Learning Theory
The math and proofs behind ML - for those who want to really understand
🔗 https://lnkd.in/gFBCyuCU

6️⃣ CS234 – Reinforcement Learning
MDPs, Q-learning, policy gradients, and agents that learn by doing
🔗 https://lnkd.in/gVvz87Hc

7️⃣ CS224U – Natural Language Understanding
Semantic analysis, dialogue systems, language meaning
🔗 https://lnkd.in/gRsWMfA6

8️⃣ CS224N – NLP with Deep Learning
Transformers, embeddings, and modern NLP architectures
🔗 https://lnkd.in/gQdFAZXM

♾️ New to AI? Start here:
 https://lnkd.in/gTQyc_fi

🎯 Why this curriculum matters:
• Not just tools, real understanding
• Theory and application from world-class professors
• Build instead of just using
• Go deeper than 95% of "AI experts"

You're one click away from learning real AI.
Which course are you starting with?

♻️ Repost to help someone learn AI the right way
➕ Follow me, Sairam, for AI engineering that works.

Join 26k+ learners from Google, Meta, Netflix, and other top tech companies and level up your AI skills for free at Gradient Ascent: https://lnkd.in/gZbZAeQW

This list will save you 100+ hours of research
19 courses on Agents, RAG, and real-world LLMs:

1. Neural Networks Zero to Hero:
 https://lnkd.in/gPy8PCSg
2. Language Modeling From Scratch:
 https://lnkd.in/g84KRW96
3. HuggingFace's Agent Course:
 https://lnkd.in/gmTftTXV
4. MCP with Anthropic:
 https://lnkd.in/geffcwdq
5. Building Vector Databases with Pinecone:
 https://lnkd.in/gCS4sd7Y
6. Vector Databases from Embeddings to Apps: 
https://lnkd.in/gm9HR6_2
7. Agent Memory: 
https://lnkd.in/gNFpC542
8. Deep Learning AI's RAG course:
 https://lnkd.in/gy7HjASS
9. Building and Evaluating RAG apps:
 https://lnkd.in/g2qC9-mh
10. Building Browser Agents:
 https://lnkd.in/gsMmCifQ
11. LLMOps: 
https://lnkd.in/g7bHU37w
12. Evaluating AI Agents:
 https://lnkd.in/gHJtwF5s
13. Computer Use with Anthropic:
 https://lnkd.in/gMUWg7Fa
14. Multi-Agent Use: 
https://lnkd.in/gU9DY9kj
15. Improving LLM Accuracy:
 https://lnkd.in/gsE-4FvY
16. Agent Design Patterns:
 https://lnkd.in/gzKvx5A4
17. Multi Agent Systems:
 https://lnkd.in/gUayts9s
18. Berkeley Agent MOOC:
 https://lnkd.in/gqyKWE3A
19. Berkeley Advanced Agents MOOC:
 https://lnkd.in/gydt98kW

Bonus- Learn the basics first:
 https://lnkd.in/gTQyc_fi

Which ones would you add to this list?

♻️ Repost this to save someone 100+ hours of research
➕ Follow me, Sairam, for more hands-on AI resources
---
Join 26k+ learners from Google, Meta, Netflix, and other top tech companies and level up your AI skills for free at Gradient Ascent: https://lnkd.in/gZbZAeQW

IBM’s 34 free videos teach AI properly.
From fundamentals to deployment.

📚 Here's what's covered in this playlist:

1️⃣ Core AI Architectures
→ CNNs, GANs, Transformers, LSTMs, Autoencoders
→ How each works and when to use them

2️⃣ Language Models
→ LLMs, Zero Shot Reasoning
→ NLP vs NLU vs NLG
→ Build RAG systems with clarity

3️⃣ Training & Optimization
→ Backpropagation, gradient descent
→ Overfitting vs underfitting
→ Includes federated learning for privacy-aware systems

4️⃣ ML Fundamentals
→ Supervised vs unsupervised learning
→ Time series, Monte Carlo
→ Core patterns to solve real-world ML problems

5️⃣ MLOps
→ PyTorch inference, embedded + edge AI
→ Deploy and scale AI systems

6️⃣ Agentic AI
→ AI Agents, LRMs, LLM as a Judge
→ Covers AI bias & trust

🎯 Why this matters:
• Debug ML issues with confidence
• Build real AI, not just prompts
• Deploy models that work

🎥 Watch the full IBM AI playlist (34 videos):
🔗 https://lnkd.in/gwypBAMy

🔗 Learn System Design for GenAI: https://lnkd.in/gqTrvsuS

♻️ Repost to help someone learn AI the right way.
➕ Follow me, Sairam, for AI engineering that works.
-----
Want to go deeper? Gradient Ascent has you covered. Join 26k+ readers from Google, Meta, Netflix, and over 160+ countries worldwide: https://lnkd.in/gZbZAeQW

The easiest way to learn the basics of AI is this: https://lnkd.in/gTQyc_fi

Stanford's Transformer & LLM essentials distilled into one readable guide.

Modern AI evolves weekly. 

Stanford's CME 295 course has condensed them into a handy cheatsheet for professionals who need answers, not lectures.

Key concepts covered include:
🔸 Core Transformer architecture & self-attention
🔸 Crucial attention optimizations (Sparse, Flash)
🔸 Advanced LLM prompting & fine-tuning (SFT, LoRA)
🔸 Preference tuning fundamentals (RLHF, DPO)
🔸 Model optimization techniques (MoE, Quantization)
🔸 Building with RAG, agents, & reasoning models

Skip months of reading. Master what matters now. 

Many thanks to Afshine & Sherwine Amidi for creating this valuable course and cheatsheet.

👉 The cheatsheet: https://lnkd.in/g52Y2HAx
👉 Bonus - learn the fundamentals of AI first: https://lnkd.in/gTQyc_fi

♻️ Repost to help someone land their dream job
➕ Follow me, Sairam, for more practical AI resources
—-
Want to go deeper? Gradient Ascent breaks down real-world AI systems every week. Join 26k+ readers across 160+ countries:  https://lnkd.in/gZbZAeQW

Transformer Cheat Sheet

The tech landscape just shifted. Again. 
Here are 50 resources to future-proof your career:

📹 Videos:
1. LLM Introduction: https://lnkd.in/gN8sv7Q5
2. LLMs from Scratch: https://lnkd.in/gJt-SQj2
3. Agentic AI Overview (Stanford): https://lnkd.in/gk4GKdxa
4. Building and Evaluating Agents: https://lnkd.in/ghBiVjdF
5. Building Effective Agents: https://lnkd.in/g_m78sid
6. Building Agents with MCP: https://lnkd.in/gAzBzr3W
7. Building an Agent from Scratch: https://lnkd.in/g9GR9b9F
8. Philo Agents: https://lnkd.in/gnxRq9n9

📚Books:
1. AI for the Rest of Us: https://lnkd.in/gTQyc_fi
2. Understanding Deep Learning: https://lnkd.in/g4xgvu_q
3. Building an LLM from Scratch: https://lnkd.in/g2YGbnWS
4. The LLM Engineering Handbook: https://lnkd.in/gWUT2EXe
5. AI Engineering: https://lnkd.in/gpfQSMCQ

🗂️ Repos
1. GenAI Agents: https://lnkd.in/gvfAc-9H
2. Microsoft's AI Agents for Beginners: https://lnkd.in/gK8MiVfv
3. Prompt Engineering Guide: https://lnkd.in/gJjGbxQr
4. AI Agent Papers: https://lnkd.in/gfKQ82Fs

🗺️ Guides
1. Google's Agent Whitepaper: https://lnkd.in/gFvCfbSN
2. Google's Agent Companion: https://lnkd.in/gfmCrgAH
3. Building Effective Agents by Anthropic: https://lnkd.in/gRWKANS4.
4. Claude Code Best Agentic Coding practices: https://lnkd.in/gs99zyCf
5. OpenAI's Practical Guide to Building Agents: https://lnkd.in/guRfXsFK

📜 Papers
1. ReAct: https://lnkd.in/gRBH3ZRq
2. Generative Agents: https://lnkd.in/gsDCUsWm.
3. Toolformer: https://lnkd.in/gyzrege6
4. Chain-of-Thought Prompting: https://lnkd.in/gaK5CXzD.
5. Tree of Thoughts: https://lnkd.in/gRJdv_iU.
6. Reflexion: https://lnkd.in/gGFMgjUj
7. Retrieval-Augmented Generation Survey: https://lnkd.in/gGUqkkyR.

🧑‍🏫 Courses:
1. HuggingFace's Agent Course: https://lnkd.in/gmTftTXV
2. MCP with Anthropic: https://lnkd.in/geffcwdq
3. Building Vector Databases with Pinecone: https://lnkd.in/gCS4sd7Y
4. Vector Databases from Embeddings to Apps: https://lnkd.in/gm9HR6_2
5. Agent Memory: https://lnkd.in/gNFpC542
6. Building and Evaluating RAG apps: https://lnkd.in/g2qC9-mh
7. Building Browser Agents: https://lnkd.in/gsMmCifQ
8. LLMOps: https://lnkd.in/g7bHU37w
9. Evaluating AI Agents: https://lnkd.in/gHJtwF5s
10. Computer Use with Anthropic: https://lnkd.in/gMUWg7Fa
11. Multi-Agent Use: https://lnkd.in/gU9DY9kj
12. Improving LLM Accuracy: https://lnkd.in/gsE-4FvY
13. Agent Design Patterns: https://lnkd.in/gzKvx5A4
14. Multi Agent Systems: https://lnkd.in/gUayts9s

📩 Newsletters
1. Gradient Ascent: https://lnkd.in/gZbZAeQW
2. DecodingML by Paul: https://lnkd.in/gpZPgk7J
3. Deep (Learning) Focus by Cameron: https://lnkd.in/gTUNcUVE
4. NeoSage by Shivani: https://blog.neosage.io/
5. Jam with AI by Shirin and Shantanu: https://lnkd.in/gQXJzuV8
6. Data Hustle by Sai: https://lnkd.in/gZpdTTYD

Don’t fear the agents replacing everyone else.
Learn to build them and own the advantage.

♻️ Repost to help someone master AI
➕ Follow me, Sairam, for AI engineering that works

Agents don’t improve by accident
They improve through intentional design.

To build better agents, you need structure.
Not just trial and error.

🎯 The Core Framework: 4 Adaptation Strategies
A 2x2 based on:
• What gets optimized (Agent or Tool)
• Where feedback comes from (Tool or Agent output)

🔹 A1: Adapt the Agent via Tool Feedback
• Agent uses a tool, sees failure, and updates
• Best for mechanics like APIs or SQL
• Eg: DeepRetrieval hit 65% vs 25% recall baseline

🔹 A2: Adapt the Agent via Self-Reflection
• Agent critiques its own output directly
• Best for logic, planning, and reasoning
• Eg: R1-Searcher beat GPT-4o-mini by 48%

🔹 T1: Adapt the Tool (Agent-Agnostic)
• Improve tools for any agent to use
• Flexible and transferable design
• Eg: Dense retrievers like Contriever

🔹 T2: Adapt the Tool to the Agent
• Freeze the agent, train tools for its quirks
• Tool learns to serve one fixed model
• Eg: R1-Code-Interpreter reached 72.4%

🎯 Why This Matters
As foundation models grow larger and more expensive to fine-tune, one path forward is to stop modifying the model.

Instead, train specialized tools that translate for that specific giant model.

🎯 Critical Trade-offs

🔸 Reliability vs Creativity 
↳ Training agents (A1/A2) risks catastrophic forgetting 
↳ Your coding agent might forget how to write poetry

🔸 Cost vs Control 
↳ Tool adaptation (T1/T2) is cheaper and lower-risk 
↳ But limited by the frozen agent's intelligence ceiling

🔸 Generality vs Specialization 
↳ T1 tools are robust and reusable 
↳ T2 tools are powerful but brittle to agent upgrades

The key insight: there is no single "best" strategy.
The choice depends on whether you can fine-tune the model and whether you have verifiable ground truth.

Paper 👉 https://lnkd.in/g3q-7Xuu
Repo 👉 https://lnkd.in/gjsjDDGg

Learn GenAI System Design 👉 https://lnkd.in/gqTrvsuS

Most teams building agents today are still guessing.
This framework gives you a structured way to decide what to optimize and how.

♻️ Repost to help someone building agents skip the trial and error 
➕ Follow me, Sairam, for AI from lab to production
-----
Join 25k+ readers from Google, Meta, Netflix, and over 160+ countries worldwide: https://lnkd.in/gZbZAeQW
Learn the basics first: https://lnkd.in/gTQyc_fi

Adaptation of Agentic AI

80% of ML system design interviews fail
On the same 7 question types.
Here are the traps and how to answer them right 👇

1️⃣ Predict Netflix Watch Times
Problem: New users and shows → sparse data
❌ Trap: Jumping straight to regression on watch hours
✅ Better: Use popular-show baselines, then layer personalization as data grows

2️⃣ ETA for Maps
Problem: Needs to run fast, at global scale
❌ Trap: Designing complex ML pipelines first
✅ Better: Start simple (distance ÷ speed). Add ML for traffic, incidents, freshness

3️⃣ Spotify Artist Recommendations
Problem: New artists lack data, heavy users skew results
❌ Trap: Relying only on collaborative filtering
✅ Better: Blend “users like you” with “songs like this one,” and plan for cold start

4️⃣ Detecting Bots
Problem: Bots adapt quickly, data is imbalanced
❌ Trap: Optimizing for accuracy and stopping there
✅ Better: Monitor drift, retrain often, include evolving signals

5️⃣ Fake News Detection
Problem: Text is noisy and easy to manipulate
❌ Trap: Using only text classifiers
✅ Better: Combine article content with who shares it and how it spreads

6️⃣ Instagram Feed Ranking
Problem: Ranking shifts user behavior long-term
❌ Trap: Optimizing only for engagement
✅ Better: Balance clicks with quality signals (reports, time spent, wellbeing)

7️⃣ Toxic Tweet Classification
Problem: Language is subjective and imbalanced
❌ Trap: Focusing only on a single metric (F1/accuracy)
✅ Better: Calibrate outputs, check bias, use human review for edge cases

💡 Pro Tip: These interviews aren’t about the “best model.” They test whether your design survives production.

👉 Learn ML System Design. Use FLASH40 for 40% off: https://lnkd.in/gqTrvsuS

♻️ Repost to help someone land their next AI role
➕ Follow me, Sairam, for AI engineering that works
---
Want to go deeper?
Gradient Ascent has you covered. Join 26k+ readers from Google, Meta, Netflix, and over 160+ countries worldwide: https://lnkd.in/gZbZAeQW
Learn the basics first: https://lnkd.in/gTQyc_fi

taplio

The Battle

Armand Ruiz

Sairam Sundaresan

Followers

Posts

Engagements

Likes

Performance ScoreArmand Ruiz's LinkedIn Performance

Performance ScoreSairam Sundaresan's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Armand Ruiz

Sairam Sundaresan

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);}Armand Ruiz's LinkedIn Performance

Performance ScoreSairam Sundaresan's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreArmand Ruiz's LinkedIn Performance