TikTok

Senior Machine Learning Engineer / Data Scientist

Heicoders Academy

Instructor

Gojek

Data Scientist

Data Analyst

Kuprion Inc

R&D Assistant

Fave

Growth Planner

Nanyang Technological University

Research Assistant

Xfers

Marketing Analyst

🤖 I use analytics, machine learning and deep learning to fight spam content at work. Classic ML algorithms, time series analysis, natural language processing, or computer vision are just some examples of the techniques I’ve used.

🧑‍🏭 I like learning and teaching things! I'm a chemical engineer turned data scientist. I self-learnt pretty much everything to move into tech. I'm now pursuing my Master's in Computer Science, teaching at a data science bootcamp, and sharing data science knowledge online.

Let’s connect. 
Personal site: https://www.travistang.com
Email: travistang.vh@gmail.com

Travis Tang

No BS AI/ML Content | ML Engineer with a Plot Twist 🥷30M+ Views 📝

Done-for you Audience Building & Content Writing for AI/ML Founders that frees up 100% of your time, so you can focus entirely on building your company  
  
Is your work too important to ignore? 💌 https://tally.so/r/wk69VJ  (or shoot me a DM) 👉
  
It's 2025 and we're at the peak of the attention economy  
  
Where Authenticity is king, visibility is currency and people wanna buy from people

Not faceless dehumanized brands  
  
So, they told you LinkedIn was "free marketing" you slapped some posts and called it strategy  
  
You've poured hours crafting the 'perfect' post, only to watch it sink without a trace..  
  
Then it dawned on you, your technical brilliance isn't enough to draw a crowd  
  
Hell, maybe you've even tried mimicking influencers, scrambling for any traction..  
  
Your post got 500 f*cking views... from your friends...  
  
Being ignored is soul-crushing  
  
It's not your fault tough  
  
When you started your company, you didn't sign-up for a degree in an Audience Building, Copywriting and LinkedIn strategy 

And here you are, sacrificing yet another late night for your vision  

"Am I just sabotaging my efforts?"  
  
"Is my life-work meaningless?"  
  
Getting traction feels impossible  
  
You hoped hiring some marketer would save your weekends but..  
  
Tech is a code they can't crack  
  
They can only churn out the lame, soul-sucking crap we all ignore on sight  
  
With armies of founders flooding feeds, your voice dissolves into meaningless white noise  
  
But in the process, you've wasted more of your precious time explaining 101 tech concept and rewriting low-quality drafts 

Done being ignored?  
  
Let's make the world listen 
  
I can help you with:  

✒️ No BS AI/ML Content  
⭐ LinkedIn Profile Revamp
🎯 Website Copy Optimization  

I've done it before:

↳ For founders backed by YC, a16z, Sequoia, and Techstars
↳ For bootstrappers in LLMs, ML, Vector DBs, Data Analytics, and Engineering
↳ For AI and Machine Learning leaders at Google, Microsoft, Airbnb, and YouTube

If they could rise above the noise, so can you

Yeah, even if you're an introvert with a "boring" niche

Still with me?  
  
Amazing—You're amazing!  
  
Cheers,  
Paolo

My booking link: 💌 https://tally.so/r/wLW0oz (or shoot me a DM) 👉

Paolo Perrone

NVIDIA just made their best AI models free.

99% of developers have no idea.

While you burn cash on OpenAI credits...
NVIDIA open-sourced Nemotron.
Their entire agentic AI family 🤯

Free Models That Actually Ship:
• Nano: Runs on your laptop
• Super: Single GPU beast
• Ultra: Multi-GPU monster

What They Do:
→ Graduate-level reasoning
→ Math without hallucinations
→ Code that compiles
→ Visual reasoning that works
→ Tool calling. No loops.

The genius part:

They published EVERYTHING.
Models. Datasets. Techniques.
Actually open source.

Game Changers:
• 6x faster on new hardware
• No API limits
• Runs on a Raspberry Pi
• Commercialize it. No strings.

The uncomfortable truth:

You're waiting for GPT-5.
Smart teams ship with Nemotron.

Who's using it:
- Accenture
- CrowdStrike  
- ServiceNow
- Deloitte

The kicker:

Download from Hugging Face. Now.
Deploy tomorrow.
Zero license fees.

Forever.

They built enterprise NIM too.
(That costs money. Models don't.)

Still paying for API credits? 

♻️ Repost before OpenAI charges per thought

No BS AI/ML Content | ML Engineer with a Plot Twist 🥷50M+ Views 📝

Sebastian Raschka, PhD just exposed the biggest secret in AI.

After dissecting 9 flagship LLMs, he found something nobody wants to admit:

We haven't invented a new architecture in 7 years.

The evidence is damning:
• DeepSeek V3: 671B parameters, still a transformer
• Kimi K2: 1 TRILLION parameters, still a transformer
• GPT-OSS: OpenAI's latest, still a transformer
• Llama 4: Meta's flagship, still a transformer

Every "breakthrough" model?
Same skeleton. Different makeup.

Here's what's actually happening:

DeepSeek V3/R1: Compressed attention (MLA) + expert routing (MoE)
Gemma 3: Sliding windows to fake longer context
SmolLM3: Removed position embeddings entirely (!)
Kimi K2: Just made DeepSeek bigger

The brutal reality:

• No new architectures since "Attention is All You Need" (2017)
• Every innovation = efficiency hack
• We're polishing, not inventing

671 billion parameters.
1 trillion parameters.
Same. Damn. Architecture.

While everyone's chasing parameter counts, Sebastian mapped the actual mutations:
• MoE: Use 37B of your 671B parameters
• Sliding attention: Pretend you have infinite context
• MLA: Compress your KV cache and pray

This isn't innovation.
It's optimization theater.

The most honest technical analysis you'll read this year.

Read it. Then ask yourself:
If we can't innovate past transformers, what happens when we hit the parameter wall?

♻️ Repost if you appreciate actual technical depth over hype

AI agent bootcamps charge $2,000 for this. You're getting it for free.

1️⃣ Design (6+ hours)
→ Langflow docs: docs.langflow.org
→ Travel planning agent template: https://lnkd.in/gmbWX5Mg
→ Agentic design patterns by Andrew Ng: 

2️⃣ Build (22+ hours)
→ LangGraph course: https://lnkd.in/g7raxjpC
→ LangSmith (modules 1-3): smith.langchain.com
→ Evaluation: https://lnkd.in/ghE-8EwA
→ Prompt engineering: 4 Methods of Prompt Engineering

3️⃣ Deploy (35+ hours)
→ LangChain Academy (modules 1-6): academy.langchain.com
→ MLOps specialization: https://lnkd.in/gmeb53UP
→ System design fundamentals: YouTube playlist
→ Distilabel for synthetic data: https://lnkd.in/gd_jJaJE

63 hours. $0.
The $2,000 bootcamp teaches the same stack with better marketing.

💾 Save this. It's the curriculum.
♻️ Repost to save someone $2,000.

Wanna build AI agents? Stop watching random tutorials.

Here's the actual learning path:

1️⃣ Design (6+ hours)
→ Langflow intro: langflow.org
→ Multi-agent workflow patterns
→ Build: Travel planning multi-agent system

2️⃣ Build (22+ hours)
→ LangGraph course: https://lnkd.in/g7raxjpC
→ LangSmith for tracing: smith.langchain.com
→ Evaluation metrics and prompt engineering

3️⃣ Deploy (35+ hours)
→ LangChain Academy: academy.langchain.com
→ MLOps fundamentals
→ LangGraph Cloud deployment

Total: 63 hours. If you follow the order.
6 months if you don't.

💾 Save this.  Start at Phase 1.
♻️ Repost if you've ever jumped straight to deployment and regretted it.

90,000 AI papers published on Arxiv.

Ilya Sutskever says you only need 27 to understand 90% of what matters.

I just went through his list. Here's what shocked me:

What's NOT on the list:
❌ No RAG papers (everyone's obsessed with retrieval)
❌ No prompt engineering papers (despite the hype)
❌ No LangChain or agent frameworks
❌ Nothing about fine-tuning optimization
❌ Zero papers on "emergent abilities"

What IS on the list:
✅ "Attention is All You Need" (obviously)
✅ "The Unreasonable Effectiveness of RNNs" (2015, still matters)
✅ "Scaling Laws for Neural Language Models" (this is the one)
✅ "Kolmogorov Complexity and Algorithmic Randomness" (wait, what?)

The pattern I noticed:

1️⃣ Ilya's not telling you to read papers about using AI.
He's telling you to read papers about how intelligence works.
Three categories dominate his list:
Fundamentals (the architecture breakthroughs)
Attention mechanisms
Residual connections
LSTMs and sequence modeling

2️⃣ Scaling (why bigger actually works)
Scaling laws
Pipeline parallelism
Training dynamics

3️⃣ Compression (the theory underneath)
Minimum description length
Kolmogorov complexity
Information theory

Here's the controversial take:
If you're only reading papers from 2023-2024, you're studying tactics.

Ilya's list is strategy:

"Attention is All You Need" is from 2017.
"Understanding LSTM Networks" is from 2015.
"Deep Residual Learning" is from 2015.
The foundations haven't changed.
We're just scaling them.

The most surprising paper on the list:

"The First Law of Complexodynamics"
It's not even about neural networks.
It's about how complex systems evolve.

This is what separates engineers who use AI from engineers who understand AI.

If you only read 5:

1) Attention is All You Need (transformers)
2) Scaling Laws for Neural Language Models (why GPT works)
3) Understanding LSTM Networks (sequence modeling)
4) Deep Residual Learning (why deep networks train)
5) Kolmogorov Complexity (the theory of intelligence)

♻️ Repost to save someone from FOMO-reading 90,000 papers

99% of AI engineers can't write a GPU kernel. 

This free CUDA course on GitHub changes that.

Zero fluff, pure metal GPU programming.

What you'll build:
• GPU kernels from scratch
• Production-ready Triton code
• Optimized matrix multiplication 
   (what makes GPT-4 fast)
• Direct cuBLAS/cuDNN integration
• Custom PyTorch CUDA extensions

The uncomfortable truth: 
Every 10x performance gain in AI comes from someone who understood CUDA.
Time to join them 👇

♻️ Repost to save someone from being a framework-only developer

🚨 Update: 800+ comments. My DMs exploded.
Can't reply to everyone. 
Grab it here: https://lnkd.in/eMQn_X-6
-------------------------------------------
Hamel's AI Evals toolkit disappears in 24 hours.

This isn't publicly available.

4 resources from the guy who:
→ Trained 2,000+ engineers on evals
→ Consulted for OpenAI, Google, Meta
→ Delivers high-value engagements for top-tier companies (ask dbt Labs why)

What you're getting:

📇 12 Evals Flashcards — the visual system for memorizing what matters
📘 The Complete Evals Guide — 144 pages of everything you need
📚 The RAG Evaluation Bible — retrieval metrics that actually work
📧 30-Day Email Course — builds evals that predict user satisfaction (free)

You won't find this on Google, GitHub, or HuggingFace.

It's in your DMs. If you act fast.

Last time I did this, 4,173 people commented. I can't DM that many again.

Comment "EVALS" + Repost ♻️ and I'll send all 4

24 hours. Then it's gone.

There's a special hell in AI/ML engineering.

It hits around Year 2.
I call it the "valley of almost useful."

You're past "Hello World" tutorials.
Not good enough for real jobs.
Stuck building demos nobody uses.

The daily torture:

Morning: "Today I'll implement that paper"
Noon: Your embeddings are somehow 7-dimensional
Evening: Stack Overflow has zero answers
Night: "Maybe I should go back to web dev"

The worst part?

You know just enough to be dangerous.
Just not enough to be valuable.

🔥 The 3 Levels of "Almost Useful"

Level 1: The Overengineering Phase
Every problem needs a neural network.
CSV parsing? Better use GPT-4.
Your hello world has 17 dependencies.

Level 2: The Imposter Syndrome Pit
Everyone else seems to ship AI products daily.
Your Twitter feed is pure pain.
"Just deployed my 5th AI startup this week!"
Meanwhile, you can't get embeddings to work.

Level 3: The Clarity Moment
You realize 90% of "AI Engineers" are just prompt engineers.
Most production AI is boring Python scripts.
The magic isn't in the model - it's in the problem.

Here's what got me through the valley:

1️⃣ Ship broken things
My first production model had 67% accuracy.
Still saved the company $200K/year.
Perfect is the enemy of deployed.

2️⃣ Join a startup, not FAANG
Startups let you touch everything.
Break things. Learn fast. Ship faster.
Big tech will teach you one tiny slice.

3️⃣ Build boring solutions
Stop chasing cutting-edge.
Most businesses need basic classification.
Not your 12-layer transformer architecture.

4️⃣ Document everything
Future you will thank present you.
Especially when debugging at 3 AM.
"Why did I set embedding_dim=7?"

5️⃣ Find your valley friends
The loneliest part is thinking you're alone.
You're not. We all went through this.
Some of us just hide it better.

The uncomfortable truth:

The valley of almost useful is necessary.
It's where you learn that AI isn't magic.
It's engineering.

And engineering is mostly debugging.
And crying.
But mostly debugging.

Still in the valley?
Good. That's where real engineers are made.

Not in bootcamps.
Not in courses.
In the trenches, shipping broken code that somehow works.

Welcome to the club.

P.S. That senior engineer who seems to know everything? They googled it 5 minutes before the meeting. We're all just better at pretending now.

♻️ Repost to help someone realize they're not alone in the valley

6 hours → 2 minutes.

One line of code.

I ran HDBSCAN on 2 million embeddings yesterday.

Snowflake quietly shipped NVIDIA's cuML and cuDF preinstalled.

No CUDA setup.

Just:

'''
import cudf.pandas
cudf.pandas.install()

import cuml
cuml.accel.install()
'''

Your entire pandas + scikit-learn pipeline → GPU. Instantly.

The benchmarks:

→ Random Forest: 5x faster
→ UMAP: 25x faster
→ TSNE: 72x faster
→ HDBSCAN: 200x faster

Same code. Same libraries. Just faster.

No rewriting. No CUDA boilerplate. No infrastructure work.

If you're running scikit-learn on Snowflake CPUs right now, you're burning money for no reason.

💾 Save this before you waste another weekend on clustering
♻️ Repost to save your team's weekends too

I pushed every agent framework until it broke. Most broke fast.

CrewAI vs LangGraph vs Agno vs PydanticAI vs Atomic Agents...
Plus SmolAgents, Mastra, and AutoGen.

The brutal truth:

Most devs hate the popular ones.

CrewAI and AutoGen? "Tricky to debug"
LangGraph? "Overly complicated abstractions"
SmolAgents' CodeAgent? "Cool idea, doesn't always work"

What actually matters:

🎯 Control vs Agency
Some frameworks let LLMs run wild (AutoGen)
Others keep them on a tight leash (Atomic Agents)
Guess which ones actually ship to production?

🛠️ Abstraction Level
High abstraction = Quick start, debugging nightmare
Low abstraction = More code, but you know what broke
Pick your poison.

💀 The Learning Curve
LangGraph: Steep as hell, but scales
CrewAI: Easy start, hits a wall
PydanticAI: Write everything yourself

The uncomfortable truth:
"AI agent frameworks are some of the worst forms of abstraction"

But here's what actually works:
• Agno for developer experience
• LangGraph for multi-agent at scale
• PydanticAI when you need control
• CrewAI if you're just prototyping

Half these frameworks won't exist in 2 years.
The winners? Whoever makes debugging possible.

Saved you 2 weeks of testing. You're welcome.
♻️ Repost if you've been burned by agent framework promises

a16z just published their AI coding report.

The math blew my mind:

AI coding costs $10k/year per developer.
But creates $100k+ in value.
That's $3 trillion in new GDP.

The entire economy of France. From a few Silicon Valley startups.

Here's what nobody's talking about:

Software development just shifted from pure people cost to people + compute cost.

Your junior developer in Bangalore?
Now competing with Claude at $10k/year.

The uncomfortable truth:

We're watching the end of cost-based outsourcing.

Not because AI replaces developers.
But because location doesn't matter when compute costs dominate.

I tested the numbers myself:

• 100k context window filled
• Claude Opus 4.1 reasoning mode
• 10k output tokens
• 3 queries/hour, 7 hours/day

= $10,500 annually

That's more than many offshore developers.

But here's the kicker:

The companies winning aren't cutting developers.
They're hiring MORE.

Why?

Because when productivity doubles, the math changes:
• 1 developer + AI = 2x output
• Cost per feature drops 50%
• Suddenly every project has positive ROI

The report shows clear winners already:

Cursor: $500M ARR in 15 months
Windsurf: $2.4B Google acquisition
Claude Code: Started a war with its own distribution channel

The pattern is obvious:

Lower barriers → More projects → More complexity → More developers needed

AWS didn't kill sysadmins.
WordPress didn't kill web developers.
AI won't kill software engineers.

It just changes what they do.

University CS programs?
Already obsolete.

Teaching algorithms when AI writes them faster?
Teaching syntax when AI handles implementation?
Teaching debugging when AI prevents most bugs?

That's like teaching calligraphy in the age of keyboards.

The real skill now:

Knowing when the AI is hallucinating.
Dragging it out of architectural holes.
Translating business needs into prompts.

The $3 trillion opportunity isn't in replacing developers.

It's in making every developer 10x more valuable.

While universities update their curriculums,
smart developers are already adapting.

Are you?

♻️ Repost if you think CS education needs a complete overhaul

The AI bubble is 17x bigger than 2000 dot-com.

I wish I was making this up.

The brutal math:

ChatGPT-3: $50M to build
ChatGPT-4: $500M to build
ChatGPT-5: $5B to build

10x cost increase. 1.2x performance improvement.

That's not a business model. That's a death spiral.

What makes this nuclear:

1️⃣ Hidden Debt Bombs
AI companies raising debt through SPVs (off the books)
Goldman found $141B in "visible" debt alone
The real number? Nobody knows

2️⃣ Zero Profitability
Every AI deployment study shows the same thing:
• Models lose money at scale
• No path to profitability exists
• Unit economics get worse, not better

3️⃣ Connected to Everything
Unlike crypto, this touches real banks
Your pension fund probably owns AI debt
When this pops, 2008 will look like practice

The uncomfortable truth:

While you're building on OpenAI's API, they're building on $5B of borrowed time.

When 2008 hit, 27 million people lost jobs.
This bubble is 4x bigger.

What are you building that survives when the funding stops?

💾 Save this before your next investor pitch
♻️ Repost if you're reconsidering that AI startup idea

I just watched Sebastian Raschka build the entire attention mechanism from scratch in 2 hours.

No abstractions. No "just use torch.nn.MultiHeadAttention."

Line-by-line Python until it clicks.

This is 's breakdown of how LLMs actually work.

Step 1️⃣ Basic self-attention (no weights) 
Shows you the raw math: how a model decides which words matter.
You see why "bank" in "river bank" gets different attention than "bank" in "savings bank."

Step 2️⃣ Add learnable parameters 
Now the model can learn what to pay attention to.
This is where it stops being a simple lookup and becomes intelligence.

Step 3️⃣ Causal masking 
The trick that makes GPT work: tokens can only attend to previous tokens.
Without this, the model cheats during training.

Step 4️⃣ Multi-head attention 
The breakthrough: run 8 attention mechanisms in parallel.
One head learns syntax. Another learns semantics. Another learns long-range dependencies.
This is why transformers work.

👉 What makes this different:

Most tutorials either:
Show you the math (incomprehensible)
Show you the PyTorch API (you learn nothing)

Sebastian does both.

You see the math implemented in code, step by step.

By the end, you understand why attention_weights = softmax(Q @ K.T / sqrt(d_k)) isn't magic.

It's just weighted averaging with learnable weights.

💡The moment it clicked for me:

Attention isn't complicated.
It's just "which words should influence this word's representation?"
Then you do that for every word.
Then you stack it 96 times (GPT-4).

That's it.

"Physical AI is the next wave."

Jensen said it. YC is betting on it.

So I started digging.

First thing I learned: the data problem is insane.

LLMs:
→ Static datasets
→ Text in, text out
→ Batch it, train it, done

Robotics:
→ LiDAR + cameras + IMUs + force sensors — all at once
→ Continuous streams, not batches
→ Petabytes per week
→ One millisecond of desync = your robot drives into a wall

The tools we use for LLMs? Useless here.

You can't throw robotics data into a Parquet file and hope for the best.

Then I found Mosaico.

Open-source. Built by engineers who spent 10+ years on autonomous driving at Ambarella.

What it solves:
→ Time-series sync across all sensors
→ Data lineage so you can debug why your robot turned left
→ Artifact versioning for perfect reproducibility

We built the data stack for language.
Now we need the data stack for movement.

I starred it. You should too: https://lnkd.in/ejUJk2UG

💾 Save for when you start exploring physical AI
♻️ Repost if you think the next 10x AI company builds robots, not chatbots

Traditional RAG: retrieve → generate → pray

Agentic RAG: retrieve → validate → retry → succeed

One is a pipeline. The other is a feedback loop.

I just built this Agentic RAG system with LangGraph + Milvus

It knows when it's wrong, and fixes itself

Here’s how:

Router: "Can I answer without retrieval?" 
✅ Yes: Skip to generation 
❌ No: Retrieve from Milvus

Grader: "Are these chunks actually relevant?" 
✅ Good: Generate answer 
❌ Bad: Rewrite query, loop back

Validator: "Does this answer match the sources?" 
 ✅ Yes: Return to user 
❌ No: Retry with different approach

It's a graph with feedback loops, not a one-shot pipeline.

The difference:
Traditional RAG commits to one path. No second chances.
Agentic RAG loops back, validates, catches mistakes, and retries until it's right.

You get a system that debugs itself before you see the output.

Why this stack:
LangGraph = control flow primitives (routing, looping, branching) 
Milvus, created by Zilliz = production-grade vector search at scale
Together: RAG that thinks instead of guesses.

Full implementation 👇

This is the RAG architecture pattern that will become standard.
♻️ Repost to save a fellow engineer from another week of RAG debugging

BFCM isn't won by the brand with the best products.

It's won by the brand that sees what's breaking first.

Most operators spend BFCM like this: 

9 AM: Campaign's crushing it, scale to $10K/day
11 AM: ROAS drops 40%, nobody knows why
2 PM: Still diagnosing
5 PM: $6K burned on a broken campaign

The strategy wasn't wrong. 

You just couldn't see the problem fast enough. 

Diagnosing "why is ROAS dropping?" right now: 

→ Open Meta Ads Manager
→ Check Shopify analytics
→ Pull GA4
→ Cross-reference Klaviyo
→ Export to spreadsheet
→ Try to connect the dots

20 minutes later you've burned $3K.

Triple Whale + Moby give you real-time answers from your complete data stack. 

👉 https://lnkd.in/gtgB8GJw

Moby has real-time access to all of it. 

Ask Moby: "Why did ROAS drop?"

Get: "iOS campaign CTR crashed 67% after creative refresh. Top SKU went out of stock an hour ago. Traffic's converting to lower-margin backup products."

Not a guess. The actual answer.

One brand caught this: 

The campaign was showing 187% ROAS. 
Moby a flagged 41% return rate. 
Actual ROAS after was returns: 73%. 

They were about to scale into a loss.

During BFCM, 20 minutes to diagnose = thousands in wasted spend. 

The brands that win can see problems in seconds, not hours. 

And yes, I am a Triple Whale partner.

👉 https://lnkd.in/gtgB8GJw

♻️ Repost if you've burned budget figuring out what went wrong

taplio

The Battle

Travis Tang

Paolo Perrone

Followers

Posts

Engagements

Likes

Performance ScorePaolo Perrone's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Travis Tang

Paolo Perrone

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);}Paolo Perrone's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScorePaolo Perrone's LinkedIn Performance