How LLMs actually learn and generate: A complete overview
Mathematical verification and checking parity by loading raw OpenAI weights directly into scratch-built architectures.
→
TANISH_SARKAR
I Build ML Systems That Are
Boring To Maintain And
Interesting To Read.
AI/ML Engineer in Progress // Specializing in GenAI architecture,MLOps, Agentic Workflows, robust data pipelines, and turning academic papers into production-grade iron // Open to AI/ML & GenAI Roles
SIM_WIDGET_V1
SYS_STATUS: ONLINE
LATENCY: 12ms
> Initializing model inference environment...
> Loading tensors... OK.
> Loading tensors... OK.
CONFIDENCE: 98.4%
MEM: 2.4GB / 16GB
01 // SYSTEM_OUTLOOK
Core Statement // 2026
"I am deeply obsessed with the gap between a machine learning model that scores 94% on an offline benchmark and one that actually survives production deployment. I don't build generic tutorial wrappers. I design clean pipelines, implement architectures from scratch to truly understand their boundaries, and build systems optimized for reliable execution over academic metric inflation."
02 // SYSTEM_CAPABILITIES
CORE COMPETENCY
Deep Learning & Transformers
CNNs, ANN models, Encoder-Decoder, Attention Mechanism, ViT, BERT-style, GPT-style, model tuning, regularization.
-> Built GPT-2 transformer from first principles with BPE tokenizer.
CORE COMPETENCY
AI-ML & Data Science
NumPy, Pandas, Scikit-learn, TensorFlow, Matplotlib, Seaborn, OpenCV, HuggingFace, PyTorch, NLP.
-> Optimized CNNs with custom regularization.
CORE COMPETENCY
Programming & Tools
Python, C++, JavaScript, SQL. Git, GitHub, AWS Certified Cloud Practitioner.
CORE COMPETENCY
Backend & Databases
REST APIs, Flask, FastAPI, PostgreSQL (pgvector), MongoDB, MySQL.
-> Architected production semantic search engine bypassing LangChain.
03 // PROJECTS
PRO_01
PYTORCH / CUDA
DECONSTRUCTING THE TRANSFORMER
Built GPT-2 entirely from scratch to understand the fundamental mechanics of self-attention. Implemented custom training loops and optimized inference paths without relying on high-level abstraction libraries.
PRO_02
[UNDER_CONSTRUCTION]
#PGVECTOR
PRODUCTION SEMANTIC SEARCH
A raw implementation using pgvector. Bypassed LangChain to reduce latency and dependencies, proving that raw SQL and bare-metal embeddings often outperform bloated frameworks in production.
PRO_03
COMPUTER VISION
CV OPTIMIZATION
Architected a highly optimized CNN for CIFAR-10. Focused on memory efficiency and parameter reduction without sacrificing accuracy. Demonstrated pruning techniques for edge deployment.
DATA_STREAM // LOSS CURVE
PRO_04
COMPUTER VISION
Real-time Facial Detection
Designed and deployed a real-time facial detection pipeline using modern computer vision techniques. Optimized for low latency inference suitable for live video streams.
PRO_05
DEEP LEARNING
Autoencoder from Scratch
Developed a neural autoencoder from the ground up to explore representation learning, latent space compression, and data reconstruction without reliance on high-level wrappers.
PRO_06
MACHINE LEARNING
Heart Disease Classification
Built an end-to-end classification pipeline for predicting heart disease presence. Deployed as a robust predictive API endpoint for reliable integration.
04 // TECHNICAL_BLOGS
Active tracking & architectural deep dives hosted at tanishverse.hashnode.dev
Diffusion Models: From Sandcastles to Stable Diffusion
Unpacking the probabilistic mechanics and step-by-step noise schedule scheduling algorithms behind generative models.
→