Biography

Sumit Yadav

AI Architect | NLP Researcher | Kaggle Competitor

Affiliation

Bachelor of Computer Engineering (Graduated 2024), Dept. Electronics and Computer Engineering, Pulchowk Campus, Tribhuvan University, Nepal

Professional Summary

Advanced Artificial Intelligence Engineer and Researcher with over 5 years of expertise in System Architecture and Natural Language Processing (NLP). I specialize in designing scalable, production-grade AI systems using Hexagonal Architecture (Ports & Adapters) and Domain-Driven Design (DDD). My technical focus spans building high-performance RAG pipelines using Vector Databases (Qdrant, ChromaDB), architecting Zero-Trust Security layers for Autonomous Agents, and solving complex Low-Resource Language challenges (creator of maiBERT).

Active Kaggle Competitor and Open Source contributor, with published research on representation geometry, LLM safety, and low-resource NLP.

Interests

• Representation Learning & Information Geometry
• AI Safety & Adversarial Robustness
• Natural Language Processing (Low-Resource Languages)
• Retrieval-Augmented Generation (RAG) Systems
• Agentic AI & Multi-Agent Systems
• Computer Vision

Technical Skills

System Architecture: Hexagonal Architecture (Ports & Adapters), Microservices, Event-Driven Architecture (SSE), REST/GraphQL APIs, Domain-Driven Design (DDD)

AI & Machine Learning: Large Language Models (Llama 3, Claude, GPT-4), RAG Pipelines, Multi-Agent Systems, Fine-tuning (PEFT/LoRA), Reinforcement Learning (RLHF), Computer Vision (YOLOv8)

Vector Databases: Qdrant, ChromaDB, Pinecone, Weaviate (Hybrid Search, HNSW Indexing)

AI Security: Model Context Protocol (MCP), Prompt Injection Defense, Guardrails AI, Zero Trust Architecture

Languages: Python (Expert), C++, C, SQL, Bash, JavaScript

Frameworks & Tools: PyTorch, TensorFlow, LangChain, LlamaIndex, vLLM, FastAPI, Docker, Kubernetes, GitHub Actions

Low-Resource NLP: Tokenizer training, Multilingual Embeddings, Unicode/Font Conversion algorithms

Professional Experience

Astha.ai, USA (Remote) — AI Architect, Security & Agentic Systems

May 2024 – Present

• Designed the core MCP-Proxy using Hexagonal Architecture (Ports & Adapters), decoupling security policy logic from the SSE transport layer
• Architected a Zero-Trust framework for Autonomous Agents where every agent interaction, tool call, and memory retrieval is verified against a strict policy engine
• Engineered a policy engine supporting v1.0 (allow/deny lists) and v2.0 (conditional logic) with Role-Based Access Control (RBAC)
• Led development of MCP-Scanner: a security analysis platform integrating 78+ attack techniques mapped to MITRE ATT&CK, leveraging Claude API for intelligent fuzzing and vulnerability enumeration

Amnil Technology Pvt. Ltd, Lalitpur — AI Engineer, RAG & Infrastructure

May 2023 – May 2024

• Developed a Retrieval-Augmented Generation system using Qdrant with Hybrid Search (Sparse + Dense vectors), improving retrieval accuracy by 35%
• Deployed and optimized open-source models (Llama 3, Mistral, Qwen) using vLLM, reducing latency by 40% via PagedAttention and KV-cache optimization
• Implemented recursive query decomposition for complex multi-hop questions
• Integrated NeMo Guardrails and built an automated “LLM-as-a-Judge” evaluation framework

Ed-Acadia, Lalitpur — Chief Data Officer

May 2022 – May 2023

• Spearheaded research into OCR and document parsing for Nepali and Maithili languages using synthetic training data for Devanagari script recognition
• Developed a large-scale semantic search system using contrastive learning for non-English educational content
• Managed a team of 3 junior data scientists, overseeing ML projects from ideation to deployment

PDSC (Plan Design Solve Create), Lalitpur — Software Coordinator

May 2022 – May 2023

• Managed technical delivery of data science consulting projects
• Conducted weekly code reviews and technical workshops for interns in Python and Machine Learning

DeepLearning.AI (Remote) — GAN Mentor

Aug 2021 – Present

• Technical mentor for the Generative Adversarial Networks (GANs) Specialization, assisting hundreds of students globally with debugging, loss functions (Minimax, Wasserstein), and architectures (DCGAN, CycleGAN)

Publications

• On the Relationship Between Representation Geometry and Generalization in Deep Neural Networks (2026, Pre-Print)
  Demonstrates that effective dimension — an unsupervised geometric metric requiring no labels — strongly predicts neural network generalization across vision and language domains. Analyzed 52 pretrained ImageNet classifiers across 13 architecture families, showing output effective dimension achieves partial r=0.75 with accuracy. Establishes bidirectional causality and cross-domain generalization to NLP encoder models and decoder-only LLMs.

• Can maiBERT Speak for Maithili? (2026, Accepted at LoResLM 2026)
  First monolingual BERT model pre-trained specifically on a custom-curated Maithili corpus. Achieved 87.02% accuracy on news classification, outperforming multilingual baselines like NepBERTa and Muril. Hugging Face Model

• SafeConstellations: Steering LLM Safety (2024, Pre-Print)
  Discovered distinct geometric patterns in embedding spaces for harmful vs. benign-but-sensitive queries. Developed a steering vector method reducing over-refusals by 73% across Claude, GPT-4o, and LLaMA without compromising safety.

• Revolutionizing Currency Security with YOLOv8 (2024, J Bus Econo Stud)
  Applied YOLOv8 to counterfeit Nepali banknote detection, achieving a True Positive Recall of 0.986 on custom dataset under various lighting conditions.

• Support Vectors are a Better Way of Text Classification for Imbalanced Data (2023)
  Demonstrated that optimized SVMs with TF-IDF often outperform Deep Learning for highly imbalanced datasets with 100+ classes.

• Machine Learning Analysis of Tirhuta Lipi (2023)

Key Engineering Projects

• SAFE-MCP Security Framework: Core contributor to Security Analysis Framework for MCP. Authored detection rules for Server Enumeration (SAFE-T1601), Tool-Chaining Pivots (SAFE-T1703), and Multimodal Prompt Injection (SAFE-T1110).
• Agents.ai & Semantic Router: Built an intelligent routing system using ChromaDB for semantic matching of user queries to expert agents.
• Nepali Chat with Doc: Full-stack RAG chatbot optimized for Nepali with real-time Preeti-to-Unicode font conversion for legacy government documents.
• IRB Robotic Arm: Image Recognition Based robotics arm for medical assistance (UN SDG3) using TensorFlow and Arduino.
• maiBERT TF: Open-sourced the first TensorFlow-based BERT model pre-trained for Maithili language.
• Maithili Lipi: Classification of Tirhuta script, foundational OCR for low-resource languages.
• Unsupervised Models: VAE, GAN, C-GAN, AC-GAN, DC-GAN implementations for latent space research.
• NEPSE Simple: Nepal stock market data platform with web scraping, automation, and Telegram Bot.
• Nepali Language Tools: Devanagari Classifier, Nepali Sentiment Classifier, Nepali OCR, Nepali Poem Generator

Competitive ML & Community

• Active Kaggle Competitor in NLP and Computer Vision challenges
• Founder of NPL Coders, organizing national-level data science hackathons
• Google Scholar Profile

Education

Pulchowk Engineering College, IOE (Tribhuvan University) — Kathmandu, Nepal
Bachelor of Computer Engineering, 2019 – 2024

Major Project: “Evaluating Auto-Encoder Transformer Language Model for Maithili Text Classification” (Foundation for the maiBERT paper)

Relevant Coursework: Artificial Intelligence, Big Data Technologies, Distributed Systems, Network Security, Compiler Design

Honors and Awards

• Winner, GritFeat AI Hackathon 2023 — 'SWIFT': Wearable device with AI for elderly fall detection (Accuracy: 0.79)
• 2x 1st Runner Up, Locus Dataverse 2023 & 2022 — NLP model for scientific abstract classification
• Winner, Best AI Project, DELTA 3.0 — Nepali Harvest: Crop disease prediction for local farmers
• Winner, IT-Meet Image Challenge 2022 — Computer Vision model for ballot paper counting
• Second Place, Docsumo DataRush 2022
• Winner, LogPoint Capture The Flag 2022 — Binary exploitation and forensics
• AI and Robotics Member — RAN
• Joint Secretary — NTBNS

Courses

• GAN Specialization (Coursera / DeepLearning.AI)
• Deep Learning Specialization (Coursera / DeepLearning.AI)
• Deep Learning with TensorFlow (EDX)
• Machine Learning (Coursera)

Lab Files

• Introduction to UNIX Commands and Shell Programming
• Lab Report on A Genetics Algorithm
• Lab Report on An Artificial Neural Network
• Lab Report on Constraint Programming
• Lab Report on Propositional Logic
• Familiarization with basic CT/DT functions

see Publications.

Contact

rockerritesh4@gmail.com | +977-9819856148 | LinkedIn | GitHub | Google Scholar

Find my CV > HERE.

Excuse me, but this, this is just a piece of paper, If I’m going to be worthy of this institution, I will show you in action. –>Tom and Jerry(Kayla)