Get live statistics and analysis of Siddharth's profile on X / Twitter

It took me almost 2 years to reach where I am today You can not become an ML engineer in barely 3 months or 6 months. This is a field that requires time, efforts, reiterations and most importantly a lot of patience. Here is your 2-year Q-wise Roadmap to become an MLE 🧵(1/n)

Don't call yourself a Machine Learning Engineer if you haven't implemented this paper !

Roles I am confident to apply for now - > Data Analyst > Machine Learning Engineer > Python Backend engineer > mlops engineer ( mostly by end of this year ) Roles for which I ain't confident enough and still lack few of the skills but are in my target zone - > Data engineer > Gen AI engineer > AI engineer > Applied AI engineer Skills I still need to develop for these roles - > Databricks > Terraform > LangGraph and Langchain > Agentic RAG > System Design > Agent automation tools > Ai agentic framework > LLMops I might have missed few especially for Applied ai one cu it's one of the most skilled roles

Text to SQL conversion automation is still a big task and there are very few good open source models for this task. Let's breakdown this - >Text to SQL conversion models are basically nothing but encoder-decoder models with a multi-attention layer and a schema linking layer in between. > The encoder processes both the user query and the database schema and generates contextual embeddings ( relation-aware encoding ) >Through schema linking, tokens in the query are aligned with corresponding schema entities >The schema aware attention mechanism then allows the model to focus on relevant parts of the schema during decoding > The decoder sequentially produces SQL tokens ( constraint-based decoding ) Where do these models lack ? - >Most of the available models lack complex queries in training data itself and hence perform poorly on cross-domain or looped queries. > Language and query requirement is not always correct from a normal user. Even spelling mistakes leads to wrong entry and causes issues during retrieval hence prompting is an important part in this task. I personally worked on this in detail when I was making an end to end project, even made a synthetic data and tried training my own SLM but failed miserably and then went on to use an Open-source model. If you want to deep dive into this, i would recommend reading these research papers first - >LLM Enhanced Text-to-SQL Generation >Next-Generation Database Interfaces: >Text-to-SQL Parsing: Concepts and Methods >RASAT: Integrating Relational Structures into Pretrained Seq2Seq Model

To all the folks learning ML, just take this as an inspiration and a learning. He is taking his time to do everything slowly and not rushing through concepts. There are various topics like MGD which we don't need to use in models from scratch every time but doing it in practice helps to develops intuitions, for ex - > SGD > FBGD > Autograd engine > Layernorm > Softmax > PCA > Backpropogation > VAE .... and many more. Its not at all compulsory but it helps you to be a better ML engineer for sure !!

A lot of beginner ML folks are confused about which projects to do. The IRIS, House Prediction, Fraud detection are outdated and don't have much value. Here are 3 intermediate level Classical ML projects that will boost your understanding- 1️⃣Buyer Persona Design Involves concepts like: > Feature Engineering > Dimensionality reduction > Clustering using DBSCAN, Kmeans 2️⃣Custom Kernel for SVM on graph structured data Involves concepts like: > Graph kernels > Support Vector Machine > Scratch coding 3️⃣Voting Ensemble for Customer Churn Prediction Involves concept like: > Multiple ML algos >Combining multiple models > Soft and Hard voting

If you want to Learn MLops quick and easy without much of hustle, a great way to start is ZenML. It is an open source framework that helps you deploy your models quickly with just a basic python library. The ZenML repo has complete guide, code walkthrough, pipeline and architecture explanation. Attaching the link to the repo below⬇️ PS- I learned about it through an article last year and it was pretty cool !!

I started with these 2 python libraries for web scraping: - BeautifulSoup - Scrapy A prof asked me to scrape some research data and stuff some time back and it was then when i found out some other good scraping libraries in python - > scholarly > newspaper3k > goose3 > selectorlib > mechanicalsoup > apify > httpcrawler > trafilatura If you know any other library which might have come to your use or is website specific like scholarly, do attach it in comments !!

Attention is only used in NLP and not in computer vision ? I saw a lot of such comments under one of my posts of why one needs to know about attention if he/she is not dealing with NLP ! To break it down, the concept of self-attention works pretty similar in ViTs like how it does in transformers for NLP, hence self-attention is an imp type of attention mechanism prevalent in CV. If you are willing to deep dive into this topic, i would recommend reading the following research papers for a better understanding of this concept - > Attention Augmented Convolutional Networks > Scaling Local Self-Attention for Visual Backbones > Criss-Cross Attention for Semantic Segmentation using CCNet > End-to-End Object Detection with Transformers I am thinking to post a detailed thread on types of attention and how they work in general, lmk if you need it !!

This was one of the best ML freelance projects I have ever worked on. It combines multiple features vectors into a single large vector embedding for match-making tasks. It contains so many concepts - > YOLO > Grabcut Mask > Color histograms > SIFT ( ORB as alternate ) > Contours > CNN > BoVW model > kmeans > normalization > Standard scaler > PCA > Cosine similarity Attaching the ML architecture for better reference

If you are looking to read about Auto Encoders or Encoder-Decoder framework and architecture, I would highly suggest reading out these 4 non-general research papers: > Auto-Encoders for Projection > Denoising Auto-Encoders > Encoder-Decoder Framework for NLP to SQL queries > Masked Auto-Encoders

Don't call yourself a Machine Learning Engineer if you haven't implemented this paper !

It took me almost 2 years to reach where I am today You can not become an ML engineer in barely 3 months or 6 months. This is a field that requires time, efforts, reiterations and most importantly a lot of patience. Here is your 2-year Q-wise Roadmap to become an MLE 🧵(1/n)

Text to SQL conversion automation is still a big task and there are very few good open source models for this task. Let's breakdown this - >Text to SQL conversion models are basically nothing but encoder-decoder models with a multi-attention layer and a schema linking layer in between. > The encoder processes both the user query and the database schema and generates contextual embeddings ( relation-aware encoding ) >Through schema linking, tokens in the query are aligned with corresponding schema entities >The schema aware attention mechanism then allows the model to focus on relevant parts of the schema during decoding > The decoder sequentially produces SQL tokens ( constraint-based decoding ) Where do these models lack ? - >Most of the available models lack complex queries in training data itself and hence perform poorly on cross-domain or looped queries. > Language and query requirement is not always correct from a normal user. Even spelling mistakes leads to wrong entry and causes issues during retrieval hence prompting is an important part in this task. I personally worked on this in detail when I was making an end to end project, even made a synthetic data and tried training my own SLM but failed miserably and then went on to use an Open-source model. If you want to deep dive into this, i would recommend reading these research papers first - >LLM Enhanced Text-to-SQL Generation >Next-Generation Database Interfaces: >Text-to-SQL Parsing: Concepts and Methods >RASAT: Integrating Relational Structures into Pretrained Seq2Seq Model

I took my first premium and its about to end in 2 days. In almost a month and a week I reached- > 1 million impressions > Gained 1300+ followers And all of this without shitposting. Just posted about research, niche-ML topics, some resources here and there once in a while and tbh I loved writing those ML breakdown posts and I want to convert them into blogs but time is an issue. I don't know if I will buy premium again this month but to everyone who liked my content, appreciated it, really thankful to all of you. And if it helped you in any way or you learnt something new, my purpose is solved. What I realised this month was - > X was a whole new experience for me. > Blue tick did help me in some way >My content reached out to a lot of big accounts and some founders/investors reached out to me and I am just grateful for that. >I spent some more time than i should have on this platform which was the downside. My content frequency from hereon might decrease but I will make sure to make it much better and come up with good posts. Much love to everyone❤️

Roles I am confident to apply for now - > Data Analyst > Machine Learning Engineer > Python Backend engineer > mlops engineer ( mostly by end of this year ) Roles for which I ain't confident enough and still lack few of the skills but are in my target zone - > Data engineer > Gen AI engineer > AI engineer > Applied AI engineer Skills I still need to develop for these roles - > Databricks > Terraform > LangGraph and Langchain > Agentic RAG > System Design > Agent automation tools > Ai agentic framework > LLMops I might have missed few especially for Applied ai one cu it's one of the most skilled roles

Let's analyze this viral post- >She gets into the Auto-Rickshaw and starts moving at a speed U. >Traffic hops in after some time, the driver applies brakes and the speed decreases and auto-rickshaw starts moving at a slower speed with pauses. > Traffic goes away, driver accelerates and gets back the original speed. This is pretty similar to how Auto-Encoders work- > The initial speed U can be treated as the input data. > The traffic part where brakes are applied can be considered as the Encoder part of Auto-encoders where compression happens and unnecessary speed is removed, i.e- useful data is transformed into a reduced latent vector. >The final part where the driver accelerates is the Decoder part of Auto-Encoders where it gains back its original speed, i.e- it decompresses the latent vector back into the original data. Auto-Rickshaw helps to travel but how does this help? It mainly helps in: >Dimensionality reduction >Noise removal >Feature learning >Anomaly detection Auto-Rickshaws are used on roads but where are Auto-encoders used ? There are various ML domains like VC, NLP, Gen Ai where Auto-encoders take up different forms and form the base of multiple architectures: >Convolutional Autoencoders >Variational Autoencoders >Recurrent Autoencoders

A lot of beginner ML folks are confused about which projects to do. The IRIS, House Prediction, Fraud detection are outdated and don't have much value. Here are 3 intermediate level Classical ML projects that will boost your understanding- 1️⃣Buyer Persona Design Involves concepts like: > Feature Engineering > Dimensionality reduction > Clustering using DBSCAN, Kmeans 2️⃣Custom Kernel for SVM on graph structured data Involves concepts like: > Graph kernels > Support Vector Machine > Scratch coding 3️⃣Voting Ensemble for Customer Churn Prediction Involves concept like: > Multiple ML algos >Combining multiple models > Soft and Hard voting

Federated Learning is an ML domain that will get a huge boost in upcoming years but at the same time might saturate quickly. Let's breakdown this in very simple terms ( I haven't completed FL yet but sharing of what i know and understood up till now ) - > You might have heard of the concept of Decentralization in Web3, Federated learning is also a similar concept in ML but here the Model Training is Decentralized. > Only the model parameters are passed and the data remains safe on the local server, hence it helps multiple orgs to train a single model without directly exchanging their data. Workflow: >A central server is hosted which contains the global model >There are multiple clients around this main server where models train locally > The updated models are sent to the main server > Now all the models are combined using an aggregation method called as Federated Averaging. > This global model is then used by clients or users again. Suggested Reads: >Advances and Open Problems in Federated Learning >Privacy-Preserving Federated Learning Using Homomorphic Encryption >FedTP: Federated Learning by Transformer Personalization >Federated learning: Applications, challenges and future directions

To all the folks learning ML, just take this as an inspiration and a learning. He is taking his time to do everything slowly and not rushing through concepts. There are various topics like MGD which we don't need to use in models from scratch every time but doing it in practice helps to develops intuitions, for ex - > SGD > FBGD > Autograd engine > Layernorm > Softmax > PCA > Backpropogation > VAE .... and many more. Its not at all compulsory but it helps you to be a better ML engineer for sure !!

Attention is only used in NLP and not in computer vision ? I saw a lot of such comments under one of my posts of why one needs to know about attention if he/she is not dealing with NLP ! To break it down, the concept of self-attention works pretty similar in ViTs like how it does in transformers for NLP, hence self-attention is an imp type of attention mechanism prevalent in CV. If you are willing to deep dive into this topic, i would recommend reading the following research papers for a better understanding of this concept - > Attention Augmented Convolutional Networks > Scaling Local Self-Attention for Visual Backbones > Criss-Cross Attention for Semantic Segmentation using CCNet > End-to-End Object Detection with Transformers I am thinking to post a detailed thread on types of attention and how they work in general, lmk if you need it !!

This was one of the best ML freelance projects I have ever worked on. It combines multiple features vectors into a single large vector embedding for match-making tasks. It contains so many concepts - > YOLO > Grabcut Mask > Color histograms > SIFT ( ORB as alternate ) > Contours > CNN > BoVW model > kmeans > normalization > Standard scaler > PCA > Cosine similarity Attaching the ML architecture for better reference