Skip to content
View Subhrajyoti8520's full-sized avatar
🎯
Focusing
🎯
Focusing
7 followers · 22 following

Block or report Subhrajyoti8520

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Subhrajyoti8520/README.md

Hi there, I'm SUBHRAJYOTI SOUMYADARSAN 👋

AI Engineer | Agentic Systems, Enterprise RAG & Computer Vision | LLMOps & Async Backends (FastAPI, Docker)

Typing SVG

  • 🔭 Currently Engineering:
    • Agentic Systems Engineering • Resilient multi-agent systems featuring tripartite consensus engines (PyTorch DNN + GGUF/Ollama + RAG), dual ReAct/deterministic execution, and vector-space telemetry
    • Enterprise RAG & MLOps • Sub-10s latency pipelines combining query rewriting, two-stage retrieval (BGE bi-encoders + BAAI cross-encoders), mathematical sandboxes, and automated Ragas evaluation
    • LLM Fine-Tuning, Model Adaptation & Efficiency • QLoRA/LoRA fine-tuning pipelines (HF PEFT vs. Unsloth Triton acceleration) with 4-bit GGUF quantization cutting model footprints by up to 69%
    • Applied Computer Vision • Deep learning vision pipelines covering custom convolutional architectures, transfer learning, semantic segmentation, and automated dataset sanitization/scraping workflows
    • Production AI Backend • Asynchronous FastAPI microservices, strict Pydantic validation, multi-service Docker Compose orchestrations, and automated CI/CD via GitHub Actions
  • 💼 Focus Areas: Autonomous Multi-Agent Consensus • Two-Stage Enterprise RAG • Computer Vision (Classification & Segmentation) • LLM Evaluation (Ragas / LLM-as-a-Judge) • PEFT & Model Quantization (GGUF/Ollama) • Async FastAPI Architectures • End-to-End Containerized LLMOps (Docker & CI/CD)
  • 📫 Contact: LinkedIn • Email

🛠️ Languages & Core Technologies

Languages & Backend Infrastructure:
Python FastAPI Docker SQLite Pydantic GitHub Actions Ruff

AI, Machine Learning & LLM Systems:
PyTorch Hugging Face ChromaDB Unsloth Qdrant Ragas LiteLLM Weights & Biases Streamlit Gradio Ollama GGUF

Operating Systems & Scripting:
Linux Ubuntu Bash

⚡ Systems Architecture & Technical Competencies

Engineering Pillar Key Implementations & Patterns
Multi-Agent AI Systems Fault-tolerant consensus engines (DNN + local GGUF + RAG), dual ReAct & deterministic execution, BGE bi/cross-encoders, tool dispatch.
Enterprise RAG & Retrieval Two-stage architectures (Recall + Rerank), Cross-Encoders (BAAI), custom mathematical evaluation sandboxes, vector indexing (ChromaDB, Qdrant).
Evaluation & Quality Benchmarking Ragas framework, automated LLM-as-a-Judge test suites, faithfulness & context precision scoring.
Model Adaptation & PEFT LoRA/QLoRA instruction fine-tuning, Hugging Face PEFT vs. Unsloth Triton acceleration, loss tracking.
AI Backend & Infrastructure Asynchronous FastAPI microservices, Pydantic validation, multi-stage Docker builds, automated CI/CD via GitHub Actions (linting, image build verification), hybrid persistence (vector & relational).
MLOps & LLMOps Operations Automated CI/CD pipelines, containerized multi-service orchestrations (Docker Compose), local model runtime serving (Ollama/GGUF), telemetry & vector-space monitoring dashboards.

🚀 Highlighted Engineering Projects

Project Focus Area Architectural Highlights
autonomous-multiagent-engine Multi-Agent AI Systems Autonomous market-arbitrage engine featuring a fault-tolerant Tripartite Consensus Engine (RAG with BGE bi/cross-encoders, local GGUF LLMs, PyTorch DNN), dual deterministic/ReAct planning, and real-time alerts.
enterprise-rag-evaluation Production RAG & Backend LLMOps Production-grade SEC 10-K RAG pipeline using query rewriting, two-stage retrieval (bi-encoder dense recall + local BAAI cross-encoder reranking), and a math sandbox. Delivered as Docker Compose microservices (UI + async backend) with <5s retrieval and <10s end-to-end latency.
ecommerce-price-cot-finetuning SLM Fine-Tuning & MLOps End-to-end pipeline contrasting direct regression (Llama 3.2) with a Chain-of-Thought engine (Qwen 2.5); leverages Unsloth Triton acceleration and 4-bit GGUF quantization, slashing model memory footprint by 69% (6.2 GB to 1.93 GB) for local Docker/Ollama serving
llm-transpiler Code Gen & Tool Calling Python-to-C++/Rust transpilation engine using AST grammar parsing, LLM tool execution, and benchmarked runtime speedups(120x).
pytorch-deep-learning-foundations PyTorch & NN Architecture A production-grade, end-to-end deep learning and computer vision repository covering fundamental tensor mathematics, custom neural architectures, non-linear classification, convolutional vision pipelines (TinyVGG), dynamic data augmentation, device-agnostic GPU acceleration and modular CLI-driven training engines.
deep-learning-cv-foundations Deep Learning Foundations A curated collection of modular, production-focused deep learning implementations and computer vision labs. This repository covers end-to-end workflows: automated data scraping, dataset verification, transfer learning backbones, pixel-level semantic segmentation, categorical tabular embeddings, collaborative filtering, error diagnosis, and standalone model serialization.

Pinned Loading

  1. enterprise-rag-evaluation enterprise-rag-evaluation Public

    A production-grade, two-stage Financial RAG architecture analyzing SEC 10-K reports. Features an evaluation sandbox (RAGAS) and an asynchronous, containerized FastAPI backend powered by Gemini Flas…

    Jupyter Notebook

  2. ecommerce-price-cot-finetuning ecommerce-price-cot-finetuning Public

    A comparative fine-tuning and MLOps pipeline for e-commerce price prediction, taking Llama-3.2 and Qwen-2.5 from raw training text to quantized 4-bit local Docker/Ollama deployment.

    Jupyter Notebook

  3. llm_fine_tuning_llama_3.2-3b-r1 llm_fine_tuning_llama_3.2-3b-r1 Public

    Fine-tuning Llama-3.2-3B into a reflective reasoning assistant using Unsloth 4-bit QLoRA, SFTTrainer, and ServiceNow R1-Distill data. Quantized to GGUF and served locally via Ollama.

    Jupyter Notebook

  4. deep-learning-cv-foundations deep-learning-cv-foundations Public

    Curated implementations of end-to-end Computer Vision and Deep Learning pipelines: ResNet-18 classification, U-Net semantic segmentation, tabular embeddings, collaborative filtering, and model seri…

    Jupyter Notebook

  5. pytorch-deep-learning-foundations pytorch-deep-learning-foundations Public

    Production-grade PyTorch deep learning workflow covering tensor math, non-linear classification, TinyVGG vision pipelines, dynamic augmentations, and modular CLI training engines.

    Jupyter Notebook