Sumit_Yadav

Sumit Yadav (Rocker Ritesh)

Email: [email protected] | Mobile: +977-9819856148
Portfolio: sumityadav.com.np | GitHub: github.com/rockerritesh | LinkedIn: linkedin.com/in/rockerritesh

Summary

AI Engineer specializing in natural language processing (NLP) and AI optimization, with 5+ years of experience developing production-grade AI systems. Currently architecting:

  • Multi-agent RAG systems with guardrails for secure information retrieval
  • Context-aware chatbots with post-conversation analysis capabilities
  • LLM evaluation frameworks for accuracy and reliability testing
  • MCP Server for easy and fast way to integrate Agents

Proven track record in AI/Machine Learning engineering across many NLP projects including Maithili text classification (low-resources) (0.87 accuracy) and multilingual document analysis systems. Authored 4 peer-reviewed publications and one open review paper on machine learning optimization/security and low-resource language processing.

Core Competencies:

  • AI Prompt Design • LLM Fine-tuning • Security Document Analysis
  • Technical Documentation • Cross-functional Collaboration • GRC Data Annotation

Education

Pulchowk Engineering College | Kathmandu, Nepal
Bachelor of Computer Engineering

Courses: SDNs, FinTech, Operating Systems, Data Structures, Big Data, Artificial Intelligence, Networking, Databases

Skills Summary

  • Languages: Python, C, C++, Bash
  • Online Courses: Deep Learning and GAN Specialization, Generative AI LLM, Image Understanding TensorFlow GCP
  • Tools/Modules: CI/CD, GIT, Pytorch, LangChain, LlamaIndex, Django, Streamlit, MySQL, GraphQL
  • Soft Skills: Leadership, Event Management, Writing, Public Speaking, Time Management
  • Hobbies: Walking, Meditation, Deep Think, Meta-Thinking

Experience

AI Engineer (Remote) | Astha.ai, USA

May 2025 - Present

  • Zero Trust Agentic System: Working on projects related to RAG, Agent identity and dataflow

AI Engineer (Full-time) | Amnil Technology Pvt. Ltd, Lalitpur

May 2024 - May 2025

  • Generative AI and Machine Learning Engineering: Working on projects related to RAG, Agent-based systems, recursive query, Chatbot, SQL Agent, and scheduling optimization. Built systems including Guardrails, LLM evaluation, and Report generation
  • LLM hosting, inference optimization, and API integration: Hosted different embedding models and completion models (e.g., LLaMA 3.3 3B model) on server using vLLM inference engine, ensuring efficient performance and easy API integration

Chief Data Officer (Full-time) | Ed-Acadia, Lalitpur

May 2022 - 2023

  • AI/ML Projects: Supervised projects and research related to Data Science. Worked on different DocumentsAI systems for low-resource languages

Software Coordinator (Full-time) | PDSC (Plan Design Solve Create), Lalitpur

May 2022 - 2023

  • Project Management: Supervised projects and research related to Data Science

GAN Mentor (Part-time) | DeepLearning.AI, Virtual

Aug 2021 - Present

  • Course - GAN Specialization: Helping students understand key concepts behind Unsupervised learning (GAN)

AI and Robotics Member (Part-time) | Robotics Association of Nepal, Lalitpur

2021 - Present

  • Making Robotics based systems: Conducted research and projects related to Computer Vision based on Raspberry Pi microcontroller

Publications

  1. SUPPORT VECTORS ARE A BETTER WAY OF TEXT CLASSIFICATION FOR IMBALANCED DATA - Presented a robust SVC method for text classification (100+ classes) using term-frequency vectorization, achieving superior test data results over neural networks

  2. Machine Learning Analysis of Tirhuta Lipi - Achieved 0.97 accuracy in Tirhuta Lipi character recognition using MobileNet embedding and logistic regression, with applications in translation and OCR for low-resource languages

  3. Revolutionizing Currency Security: A Yolov8-Based Approach for Automated Detection of Counterfeit Nepali Banknotes - The challenge of counterfeit currency in financial transactions requires innovative solutions for the efficient and accurate verification of banknotes. The study focuses on categorizing banknotes as fake or genuine in the context of Nepalese currency, which is crucial for maintaining economic stability, preventing financial fraud, and ensuring public trust in the monetary system. For the 180 collected samples of 1000 Nepalese rupee banknotes, the YOLOv8 algorithm achieved a true positive recall of 0.82 for the front face and 0.9863 for the back face. The mean average precision and confidence threshold achieved in the study indicate significant improvements with YOLOv8. The proposed approach demonstrates the potential for global implementation and adaptability across various hardware platforms.

  4. SafeConstellations: Steering LLM Safety to Reduce Over-Refusals Through Task-Specific Trajectory - LLMs increasingly exhibit over-refusal behavior, where safety mechanisms cause models to reject benign instructions that superficially resemble harmful content. This phenomena diminishes utility in production applications that repeatedly rely on common prompt templates or applications that frequently rely on LLMs for specific tasks (e.g. sentiment analysis, language translation). Through comprehensive evaluation, we demonstrate that LLMs still tend to refuse responses to harmful instructions when those instructions are reframed to appear as benign tasks. Our mechanistic analysis reveal that LLMs follow distinct "constellation" patterns in embedding space as representations traverse layers, with each task maintaining consistent trajectories that shift predictably between refusal and non-refusal cases. We introduce SafeConstellations, an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways. By selectively guiding model behavior only on tasks prone to over-refusal, and by preserving general model behavior, our method reduces over-refusal rates by up to 73% with minimal impact on utility-offering a principled approach to mitigating over-refusals. (August 2025)

  5. Can maiBERT Speak for Maithili? - Natural Language Understanding (NLU) for low-resource languages remains a major challenge in NLP due to the scarcity of high-quality data and language-specific models. Maithili, despite being spoken by millions, lacks adequate computational resources, limiting its inclusion in digital and AI-driven applications. To address this gap, we introducemaiBERT, a BERT-based language model pre-trained specifically for Maithili using the Masked Language Modeling (MLM) technique. Our model is trained on a newly constructed Maithili corpus and evaluated through a news classification task. In our experiments, maiBERT achieved an accuracy of 87.02%, outperforming existing regional models like NepBERTa and HindiBERT, with a 0.13% overall accuracy gain and 5-7% improvement across various classes. We have open-sourced maiBERT on Hugging Face enabling further fine-tuning for downstream tasks such as sentiment analysis and Named Entity Recognition (NER). (September 2025)

Projects

Vibe-Coder

Made an Agent that will do Streamlit and FastAPI. Tech: Agent, MCP, Claude API keys, Python, Streamlit

Retrieval Augmentation Generation System (RAG) and Intelligent Document Processing (IDP)

Developed a retrieval-augmented reality system for enhanced information access and interaction. Tech: OpenAI, Gemini, Claude API keys, Python

Nepali Chat with Doc

Implemented a chatbot for Nepali language using Devanagari and Preeti fonts. Features include Guardrails system, post-conversation analysis, and agent-based systems like SQL Agent, Excel Agent, and Reflexive Agents. Preeti to Unicode Conversion. Tech: OpenAI, Gemini, Claude API keys

Bachelor's Major Project: Evaluating Auto-Encoder Transformer Language Model for Maithili Text Classification

Established a benchmark in this language. First to create a corpus in Devanagari Maithili language, trained LLM for Maithili, and performed downstream task classification. Tech: LLM, Transformer (BERT), PyTorch, Streamlit & Big Data (April 2024)

IRB (Image Recognition Based) Robotics Arm

Research-oriented, open-source project under UN's SDG3 - Good Health & Well-Being. Tech: Python, Arduino Programming, Arduino Toolkit, TensorFlow (May 2020)

Nepali Language Projects

Developed multiple applications, including a Devanagari letter classifier using VGG16 (accuracy 0.94), a Nepali sentiment analysis model, and a simple OCR for Nepali text. Tech: Keras, Transformer, PyTorch, TF-IDF, NLTK (Past 2 Years)

Unsupervised Model

Explored the behavior of latent spaces using VAE, GAN, C-GAN, AC-GAN, and DC-GAN. Tech: Python, NumPy, TensorFlow (Sep 2021)

NEPSE Simple

Presented Nepal stock market data in a minimal environment constraint. Tech: GitHub Workflow, Automation in Scraping, WebSockets, JavaScript, RSS, XML (Since 2020)

Advanced Document and AI Systems

Designed and implemented a variety of tools, including:

  • Chat systems for Nepali and multilingual documents with Preeti-to-Unicode conversion and guardrails for improved user interaction
  • AI-powered memo creation and advanced Excel file manipulation tools
  • Contract document analysis using recursive and advanced reasoning GPT systems
  • Translation systems for Nepali documents using OCR and text conversion
  • Chat and interaction systems for image and audio data with TTS and Whisper integration

Verification and Financial Prediction Systems

Developed:

  • A face and signature verification app using VGG-based advanced face detection and liveness detection algorithms
  • A loan eligibility prediction system utilizing knowledge-based reasoning techniques

Honors and Awards

  • Winner of GritFeat AI Hackathon 2023, Locus - Feb 2023 (SWIFT is a wearable device with hardware and AI models that detect falls in elderly people with 0.7986 accuracy, resulting in immediate emergency alerts to contacts)

  • First Runner-Up of Dataverse, Locus - Jan 2023, Dataverse Solution (NLP-based problem to classify abstracts)

  • Winner of Best AI Project of Deltathon, DELTA 3.0 - Jan 2022, Nepali Harvest (Designed a portal to help farmers by predicting diseases, identifying optimal harvest times, and aiding with crop health assessment)

  • Winner of Image Challenge, IT-Meet UP KU - Sep 2022 (Training AI model to classify images of ballot papers)

  • Winner of Capture The Flag, LogPoint - Feb 2021 (Task of finding information and exploiting a binary file)

  • Runner's Up at DATARUSH by DOCSUMO - Feb 2021 (NLP-based model for classifying abstracts into classes)

Social Experience

Team of NPL Coders | Global

Sep 2023 - Present
Conducted National level Data Science Coding Competition on Kaggle and HackerRank

Joint Secretary at NTBNS Student Clubs, IOE, Pulchowk Campus | Lalitpur, Nepal

Jan 2020 - Present
Conducted technical training & Organized Nepal's largest Saraswati Puja Program

Tutor of Children In Technology - WorldLink | Nepal

Nov 2023
Educated students about risks and safety of the Internet