Tushar Raj Verma

Senior Data Scientist @TOTHENEW

Creating Machine Learning Threads.

[Tushar Raj Verma @TrajVoid]

Everything about using docker with python (Flask or Django) and ML in Production (TensorFlow Serving).

I invested my time in this course and have completed proof of concepts.

Python in containers helped me for
‣ Dockerizing applications
‣ Deploy ML models
‣ Multiple services

[Tushar Raj Verma @TrajVoid]

From knowledge distillation → knowledge bases
From tensors → tokens
From SageMaker → Bedrock
From epochs → API requests
From dimensionality errors → max-token exceeded

We’ve truly come a long way.

[Tushar Raj Verma @TrajVoid]

Amazing to see how the data science world has evolved
From fine-tuning neural networks → tuning prompts
From writing nested loops for similarity → using vector DBs
From supervised learning → giving examples in prompts
From accuracy metrics → LLM-as-a-judge

[Tushar Raj Verma @TrajVoid]

Day 1 of the Data Hack Summit 2025 was amazing, learned a lot about AI agents, in-house LLMs, and MCPs. Got to see some brilliant minds and innovative products. It was a real pleasure meeting @joshuastarmer, whose tutorials I’ve been following for years - Triple Bam!

[Tushar Raj Verma @TrajVoid]

Thought of posting today after facing issues adding 300k+ embeddings (356-dim) to a vector DB:

• Overflow & NaNs during normalization
• Clipped embeddings values to [-10,10] to fix
• Manual normalization after clipping
• Used Qdrant with DOT distance for similarity search

[Tushar Raj Verma @TrajVoid]

@capeandcode Have you checked the langchain agent's latency?

[Tushar Raj Verma @TrajVoid]

It's time to combine GraphDBs with vector databases to achieve better results in edge cases where traditional vector searches struggle to find relevant chunks.

Feels like the XGBoost approach for RAG

https://t.co/pBqO9yGNJw

[Tushar Raj Verma @TrajVoid]

Check out my design addressing the max token length prompt issue.
No more sharing the entire document repository at once in sequence with any API service.
💡 Open to suggestions and recommendations! 🔍

[Tushar Raj Verma @TrajVoid]

I worked on some fine-tuning exercises for a code generation [StarCoder] AI model last week.

Here's a blog:
https://t.co/a8JvOMOWAU

[Tushar Raj Verma @TrajVoid]

Tried this method recently in a project, it's solid. Code's way easier to read now.

Pipeline's saved pickle file is production-ready without extra test time code for preprocessing or transformations.

Even added it to my custom pipeline wrapper.

https://t.co/EqSXIdVhR9

[Santiago @svpino]

One of the biggest mistakes data scientists make:

They don't use pipelines.

Pipelines will instantly improve your data transformation and training processes.

Three advantages:

• Cleaner code
• More robust
• Easier to deploy in production

Example attached.

[Tushar Raj Verma @TrajVoid]

Here is the badge representing my achievement.

https://t.co/kf6Tuce1y2

[Tushar Raj Verma @TrajVoid]

After thorough preparation, I successfully passed the Databricks Machine Learning Associate exam.
Throughout the past year, I had been creating content and implementing most of the modules, which greatly contributed to my success.

https://t.co/jNbWRRy1Pw

[Tushar Raj Verma @TrajVoid]

For about two months, I dedicated most of my time to various @databricks courses related to Machine Learning.

Getting practical knowledge about this platform and upgrading my skills.

I have explored advanced skills that will help me create insightful content.

1/5

[Tushar Raj Verma @TrajVoid]

Here is an in-depth overview of my preparation process from last year.

https://t.co/zJYVTZXQcE

[Tushar Raj Verma @TrajVoid]

Resources and a course roadmap with the implementation platform of my previous thread related to @databricks courses.

‣ You will find here GitHub repositories specific to Machine Learning.

‣ Platform to implement those repositories and setup.

1/8
https://t.co/rfbt1gyAae

[Tushar Raj Verma @TrajVoid]

@_avichawla Great post!!
I usually skip scaling when working with XGBoost or RF as they handle features differently with their split node. In production, avoiding the extra code for scaling makes things simpler. But for ANN, scaling is crucial to prevent loss from going nan(large input)

[Tushar Raj Verma @TrajVoid]

Zero-shot learning is pretty cool and makes a big impact when used for things like images and text.

I've been working on it, and I'm still at it!

https://t.co/3licGDID8e

[Tushar Raj Verma @TrajVoid]

‣ Sentence Similarity is used in various business problems for automating the matching text.

‣ For me, sentence similarity is an inference version of OneShot Learning mainly used in image domains.

‣ Here is what I mapped a flow to compute similarity between two sentences.

[Tushar Raj Verma @TrajVoid]

Excited to share my blog post on AWS training job with XGBoost! 🚀 Dive into the world of cloud computing and machine learning.
Check it out here:
https://t.co/i6zcIBedda

[Tushar Raj Verma @TrajVoid]

@mathsppblog Solving a set of recursion problems.

[Tushar Raj Verma @TrajVoid]

@__mharrison__ @JonKrohnLearns Awesome! I've been working on XGBoost for a while. I'm wondering about my recent implementation if it's a good algorithm for training on text embeddings(bert, adaa) and numerical features to boost performance in a specific area.

[Tushar Raj Verma @TrajVoid]

Check out my design addressing the max token length prompt issue.
No more sharing the entire document repository at once in sequence with any API service.
💡 Open to suggestions and recommendations! 🔍

[Tushar Raj Verma @TrajVoid]

I will be sharing a sequence of threads related to use cases for GenAI. Please stay tuned @TrajVoid

[Tushar Raj Verma @TrajVoid]

🌐🔊 Exciting Tech Journey: From English Audio to Multilingual Voices! 🔊🌐

1/ 🎙️ Started with English audio, but the world is diverse! 🌍 I transformed English audio to text using the power of code. 🧑‍💻 Let's explore how I bridged the language gap!
https://t.co/CZKZ01p0z4

[Tushar Raj Verma @TrajVoid]

4/ ⏸️ Introducing the Art of Pauses: It's not just about words, but the rhythm of speech! 🎵 I implemented pause logic to make the conversation more natural, adding a human touch that goes beyond translation. Added Pauses with silence and timestamp logic

https://t.co/dSu8Kiyx2X