Skip to content
View sujal-maheshwari2004's full-sized avatar
8 followers · 1 following

Achievements

Achievement: Pull Sharkx2Achievement: Pair ExtraordinaireAchievement: YOLOAchievement: Quickdraw

Achievements

Achievement: Pull Sharkx2Achievement: Pair ExtraordinaireAchievement: YOLOAchievement: Quickdraw

Highlights

  • Pro

Block or report sujal-maheshwari2004

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.

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
sujal-maheshwari2004/README.md

header

Typing SVG

Portfolio LinkedIn HuggingFace Resume Email


About Me

name     : Sujal Maheshwari
degree   : B.Tech CSE (AI & Data Science) · Graphic Era University, Dehradun
location : Aligarh, Uttar Pradesh, India
status   : Final Year · 2026
seeking  : Internships · Research Roles · Freelance (AI / LLMs / Full-Stack)
exploring: Agentic AI workflows with MCP

I build production-grade LLM systems, train language models from scratch, and ship agentic infrastructure that solves real engineering problems. My work spans the full vertical — custom tokenizers and model architectures, RAG pipelines, LangGraph agent loops, MCP servers, FastAPI backends, and everything in between.


⚡ Impact at a Glance

Metric Result Context
🚀 Latency reduction 96% 7 min → 30 sec · LangGraph parallel agents
🎯 RAG accuracy 62% → 86% Repo indexing + AST chunking pipeline
🤖 LLM parameters 125M Trained entirely from scratch · Librarian series
📊 Records processed 130K+ Shopify pipeline · 5 hrs → under 45 min
👨‍🏫 Students mentored 130+ Workshops & project guidance · Graphic Geeks Club

💼 Work Experience

 🔵  AI Architect · STARTHACK.IO (Freelance)  |  Sept 2025 – Feb 2026 · Dehradun, Hybrid
  • Built parallel LangGraph agents for automated code evaluation — cut end-to-end latency from 7 min to 30 seconds
  • Architected a real-time voice-to-voice technical interview system using GPT Real-Time APIs with dynamic persona switching and low-latency audio streaming
  • Designed scalable repository indexing and code chunking strategies for a RAG pipeline, improving retrieval accuracy from ~62% to ~86%

LangGraph GPT-4o GPT Real-Time API RAG AST Azure GitHub Actions

 🟣  AI Data Scientist Intern · Basal AI (Remote)  |  May – Aug 2025 · Bengaluru
  • Fine-tuned GPT-4.1 on domain data via Azure AI Foundry — content generation turnaround from 3 days → 15 minutes at 96% baseline accuracy
  • Built a parallelised Shopify scraper — reduced product and variant compilation time from 5+ hours to under 45 minutes across 130K+ records
  • Built Teams-integrated agents using Power Automate and Copilot Studio for internal analytics workflows and automated reporting
  • Configured CI/CD pipelines with GitHub Actions for automated testing and deployment of AI inference services

GPT-4.1 Azure AI Foundry Python Power Automate Copilot Studio GitHub Actions

 🟢  LLM Engineer · Patent Work  |  Apr – Jul 2024  ·  App No: 202411035697
  • Built an NLP keyword-extraction chatbot using rule-based and statistical filtering on hospitality-domain data
  • Fine-tuned GPT-2 on proprietary corpora and implemented a RAG pipeline for document-grounded responses
  • Contributed to a patentable AI system — Namaste-Enabled Service Robot for Hospitality Industry

NLP GPT-2 RAG LLM Hospitality AI


🚀 Projects

🧠 DriftGuard

Agentic Infra · Semantic Memory Layer

Mistake-memory system for AI agents. Stores causal chains (action → feedback → outcome) in a knowledge graph and warns agents before repeating past failures — using semantic similarity, not exact matching.

Dual interface: MCP server for tool-based agents + in-process guard API for LangGraph loops. Full pytest coverage, CI via GitHub Actions.

MCP NetworkX spaCy sentence-transformers

📈 Bot Street

Simulation · HFT Environment for LLM Agents

Full algorithmic trading simulator on Apache Kafka (KRaft). Price-time priority order book, circuit breakers, sentiment engine, 15+ quant indicators (RSI, MACD, Bollinger, OFI, VaR, CVaR). An MCP server lets LLM agents trade live against bots.

Participants: Market Maker · Momentum · Mean Reversion · Random · LLM Agent

Kafka FastAPI MCP Docker

📚 Librarian Series

Personal Research · LLM From Scratch

125M parameter causal LM built end-to-end: custom 16K BPE tokenizer, GPT architecture with RoPE + RMSNorm + SwiGLU MLP, trained on WikiText-103 + TinyStories. 6.19 perplexity on validation. LoRA fine-tuned on DailyDialog. Ships with a full config-driven SFT framework.

Model Status
librarian-base-130m ✅ HF
Librarian-Instruct-130m ✅ HF
librarian-sft framework ✅ Released
librarian-base-390m 🔜

PyTorch LoRA HuggingFace

🔧 ToolStore

LLM Infra · Automatic MCP Server Builder

Plain-English tool description → semantic search + cross-encoder reranking → clone matched repos → static AST security scan → single runnable MCP server. 100% build accuracy and AST validity across 32,767 evaluated subsets.

MCP ChromaDB AST sentence-transformers

⚡ PeakPulse

AI Product · Customer Support Intelligence

3-node LangGraph pipeline — classify → route → resolve — with deterministic escalation and rule-based fallback at <5ms, $0 cost. 50 concurrent queries via asyncio.gather.

LangGraph GPT-4o-mini React

📰 NewsCheck

Pipelines · Audio → Structured Summary

YouTube link → yt-dlp audio → Whisper transcription → GPT-4 summary in 5 structured bullets. Magic-link JWT auth, per-user rate limiting, deployed on Azure VM.

Whisper FastAPI Azure VM

🏥 MediMatch

AI/ML · Medical Document Parsing

Organ transplant prioritization using LLaMA 3.2. Parses medical documents and ranks patients by criticality score using structured LLM output. Node.js/Express backend.

LLaMA 3.2 Node.js GenAI

🛠️ Tech Stack

Languages

Python JavaScript C++ SQL

AI / LLM

LangChain OpenAI PyTorch HuggingFace Whisper

Backend & Web

FastAPI React Node.js Flask Kafka

Cloud & DevOps

Azure Docker GitHub Actions Prometheus

Databases

PostgreSQL MongoDB Redis ChromaDB


📊 GitHub Stats

 

🏅 Certifications

Certification Issuer
🔷 DP-900 · Microsoft Azure Data Fundamentals Microsoft
🟠 Generative AI Specialization AWS + DeepLearning.AI
🔵 Data Analytics IBM
🟡 AWS Certified Cloud Practitioner Amazon Web Services

✦ Highlights
🏆 Top Performer · Smart India Hackathon 2024 — AI for Smart Education
📄 Patent Contributor · Namaste-Enabled Service Robot for Hospitality Industry · App No: 202411035697
🎓 Founder & Technical Team Lead · Graphic Geeks Club — mentored 130+ students via workshops and project guidance
🤗 Published on HuggingFace — 125M param LLM trained from scratch, publicly available
🔭 Currently building agentic AI workflows with MCP

footer

Let's build something intelligent together

Email  ·  LinkedIn  ·  Portfolio

Pinned Loading

  1. ToolStore ToolStore Public

    ToolStorePy -- Describe the tools you need in plain English. ToolStorePy finds the best match from a curated index, audits the code for security issues, and generates a ready-to-run MCP server. One…

    Python 1 2

  2. Bot-Street Bot-Street Public

    Full algorithmic trading simulator on Apache Kafka (KRaft). Price-time priority order book, circuit breakers, sentiment engine, 15+ quant indicators (RSI, MACD, Bollinger, OFI, VaR, CVaR). An MCP s…

    Python

  3. DriftGuard DriftGuard Public

    Semantic mistake-memory and guardrail for AI agents. Stops agents from repeating the same failures using causal graphs and semantic matching. Supports MCP, LangGraph, and in-process Python.

    Python 2

  4. librarian-base-v1 librarian-base-v1 Public

    Python 1

  5. newsCheck newsCheck Public

    NewsCheck Pipelines · Audio → Structured Summary YouTube link → yt-dlp audio → Whisper transcription → GPT-4 summary in 5 structured bullets. Magic-link JWT auth, per-user rate limiting, deployed o…

    JavaScript

  6. Vigil-ai Vigil-ai Public

    A self-healing AI reliability platform with observability, anomaly detection, and lightweight MLOps workflows including dataset/model versioning, registry-backed inference, and automated retraining.

    Python