Hire the Best NLP Engineers
in Nepal

Hi, I'm Bikash a Machine Learning Engineer who takes AI systems from prototype to production: fine-tuned LLMs, retrieval-augmented generation, inference APIs, and NLP/audio pipelines built to actually run in real applications, not just notebooks. I help businesses move from AI ideas to working software: training and fine-tuning models, building APIs, optimizing inference, integrating ML into applications, and deploying systems that are practical, reliable, and easy to maintain. What I can help you with: 1. LLM fine-tuning and retrieval-augmented generation (RAG) systems, with structured, citation-grounded outputs 2. Machine learning models for classification, prediction, scoring, and decision support Production ML workflows using Python, TensorFlow, PyTorch, Scikit-learn, FastAPI, Docker, and cloud deployment 3. NLP systems for summarization, translation, semantic search, recommendations, and text-to-speech 4. Audio AI systems, including sound classification, scoring logic, temporal smoothing, and lightweight inference APIs 5. Backend integration for ML products using REST APIs, FastAPI, Flask, databases, and cloud deployment What I've actually built: 1. Khabar AI: an end-to-end AI news platform using 36+ RSS sources, mBART summarization, English-to-Nepali translation, Qdrant semantic search, recommendations, newsletters, and Piper TTS. 2. MarketGyan: a Nepal-focused financial NLP system extending Khabar AI into Nepal Stock Exchange (NEPSE) market analysis. I built a 500-row bilingual benchmark and fine-tuned a Qwen3.5-9B LoRA adapter for structured, evidence-grounded extraction, reaching 1.000 structured-output validity and 0.796 relevance macro-F1, then deployed a retrieval-augmented generation runtime delivering cited, disclaimer-gated market answers. 3. Laughter Detection & Scoring System: a cloud-deployed audio classification API using YAMNet embeddings, TensorFlow Lite optimization, FastAPI, and Google Cloud Run. I built probability aggregation, temporal smoothing, and noise-aware thresholds, reaching 93.33% accuracy, 0.909 F1, 1.0 recall, and 0.95 balanced accuracy. 4. EEG Epilepsy Detection System: a deep learning signal-classification pipeline for EEG-based epilepsy detection, reaching 99.47% classification accuracy in academic evaluation. 5. Text-to-Speech System: a Nepali TTS project using Coqui TTS, Glow-TTS, HiFi-GAN, Gradio, and Librosa, reaching 4.07/5.0 MOS with about 2-second CPU inference per 10 words. I'm also currently a Research Intern at Kathmandu University's Information and Language Processing Research Lab (ILPRL), working on active learning and low-resource speech ASR so I stay current with the field, not just applied delivery work. Core skills and tech stack: Machine Learning: production ML, model serving, inference APIs, ML pipelines, deep learning, model evaluation, active learning AI/NLP/LLMs: semantic search, summarization, translation, speech synthesis, audio classification, LLM fine-tuning (LoRA), retrieval-augmented generation Frameworks: Python, TensorFlow, PyTorch, Scikit-learn, Keras, Pandas, NumPy, Librosa, XLM-R, Qwen Deployment: FastAPI, Flask, REST APIs, Docker, TFLite, Google Cloud Run, vLLM, Gradio, Git, Linux Databases and search: SQL, PostgreSQL, MongoDB, SQLite, Qdrant vector search Certifications: Google Cloud Computing Foundations, Dataquest Data Scientist in Python, Dataquest Deep Learning in TensorFlow I focus on clean code, clear communication, measurable results, and AI systems that can actually be integrated into real applications. I also have fluent English proficiency, with IELTS 7.0 / CEFR C1. If you need a Machine Learning Engineer who can turn your idea into a working model, API, or deployable AI system, send me a message and I'll help you plan the best approach.

𝗔𝗜 / 𝗡𝗟𝗣 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿 | 𝗥𝗔𝗚 & 𝗟𝗟𝗠 𝗦𝘆𝘀𝘁𝗲𝗺𝘀 | 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵-𝗕𝗮𝗰𝗸𝗲𝗱 𝗔𝗜 𝗦𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝘀 I am an AI and NLP Engineer and published research paper author, specializing in Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), and applied NLP systems. My work focuses on accuracy, evaluation, and real-world deployment, particularly for domain-specific and low-resource language applications. I work at the intersection of research and production engineering—designing AI systems grounded in peer-reviewed methods, validating them with rigorous evaluation, and delivering production-ready solutions rather than experimental demos. I am the winner of the AIDEA National-Level AI Hackathon, where my project NepSAUL was recognized for technical depth and real-world impact. The project was subsequently selected for angel investment seed funding, validating its practical and commercial viability. 𝐑𝐄𝐒𝐄𝐀𝐑𝐂𝐇 & 𝐏𝐔𝐁𝐋𝐈𝐂𝐀𝐓𝐈𝐎𝐍𝐒 (𝐀𝐔𝐓𝐇𝐎𝐑) • Profanity and Offensiveness Detection in Nepali Social Media Using Bi-directional LSTM Models 21st International Conference on Natural Language Processing (ICON 2024) • Evaluating Sentence Embedding Models for Nepali Sentiment Analysis National Conference on Computer Innovation 2025 • Retrieval-Augmented Generation Framework for the Nepali Legal Domain Question Answering (Under Review) 𝐂𝐎𝐑𝐄 𝐄𝐗𝐏𝐄𝐑𝐓𝐈𝐒𝐄 𝗥𝗔𝗚 & 𝗟𝗟𝗠 𝗦𝘆𝘀𝘁𝗲𝗺𝘀 • Designed and implemented NepSAUL using 10,000+ real court case documents • Hybrid retrieval using BM25 + dense embeddings • Achieved 91% Precision@1 in low-resource legal data • Grounded generation with LLM-as-Judge and expert review 𝗡𝗮𝘁𝘂𝗿𝗮𝗹 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗣𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 • Sentiment analysis, text classification, profanity detection • Created and annotated 11,000+ real-world Nepali samples • Noisy, multilingual, domain-specific data handling 𝗠𝗟 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 & 𝗗𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 • Python, FastAPI, production inference APIs • Research to production model transition • Maintainable and evaluation-driven architectures 𝐓𝐄𝐂𝐇𝐍𝐈𝐂𝐀𝐋 𝐒𝐊𝐈𝐋𝐒 • Python, C++, JavaScript • LangChain, Hugging Face, Multilingual BERT, Bi-LSTM • BM25, FAISS, Pinecone • TensorFlow, Keras, Scikit-learn • FastAPI, Flask 𝐖𝐇𝐘 𝐂𝐋𝐈𝐄𝐍𝐓𝐒 𝐂𝐇𝐎𝐎𝐒𝐄 𝐌𝐄 • Research-backed engineering decisions • Clear and transparent communication • Production-focused system design • Honest feasibility assessment 𝐀𝐕𝐀𝐈𝐋𝐀𝐁𝐈𝐋𝐈𝐓𝐘 If you are building a high-accuracy RAG system, LLM-powered research tool, or NLP pipeline using real-world data, I am available to review requirements and propose a technically sound architecture.

I build and ship production AI systems that real users depend on, not demos. RAG pipelines, multi-agent LLM apps, fine-tuned models, and multimodal/OCR extraction, deployed to run 24/7 on Kubernetes and serverless GPU. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Top-Rated Plus | 100% Job Success | 4+ years ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Enterprise-grade AI for multinational companies and startups, including HIPAA-conscious healthcare workflows. I turn complex requirements into intelligent, production-ready applications that drive measurable results. ━━━━━━━━━━━━━━━━━━━━━ WHAT I DO BEST ━━━━━━━━━━━━━━━━━━━━━ Agentic AI & Multi-Agent Systems Custom architectures with LangGraph, CrewAI, and Model Context Protocol (MCP), including self-improving agents that learn from evaluation feedback. Built for real automation, not chatbot demos. Advanced RAG, Evaluation & Observability 10+ production RAG systems (self-RAG, adaptive retrieval), one serving hundreds of users across thousands of documents. Migrated Pinecone to Weaviate for better recall at lower cost. Every system ships with LLM-as-judge, retrieval metrics, and full tracing (LangSmith/Langfuse), so quality is measured, not guessed. LLM Fine-Tuning & Cost Optimization PEFT (LoRA/QLoRA), SFT, DPO, and instruction tuning. Fine-tuned a 7B Arabic model served on autoscaling serverless GPU, plus multimodal vision-language models. Cut client AI costs by up to 40% through open-source replacement and quantization, with no drop in performance. Multimodal & Document AI OCR and document-extraction pipelines across PDF, DOCX, PPTX, Excel, and images, with strong F1 on messy financial and clinical documents. Also built a temporal, multi-hop knowledge graph over an encrypted Postgres + Qdrant store with client-side encryption. AI Automation & Integrations Connecting LLMs to real business systems: n8n, Make (Integromat), Zapier, CRM automation (HubSpot, GoHighLevel, Airtable), Supabase backends, and Twilio/WhatsApp. AI that plugs into how your team actually works. Enterprise Backend & Scalable Infra Master-level Python (FastAPI, Flask), robust CI/CD, and multi-cloud deployment (AWS, Azure, GCP). Docker + Kubernetes with KEDA autoscaling, plus privacy-first, multi-tenant systems (E2EE, RBAC, audit logging), including HIPAA-conscious PHI handling. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ TECH STACK ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ▸ Agents & LLMs: LangChain, LlamaIndex, LangGraph, CrewAI, MCP, Hugging Face (PEFT/TRL), Ollama, TGI, vLLM ▸ Eval & Tracing: LangSmith, Langfuse, LLM-as-judge, custom eval frameworks ▸ Vector DBs: Weaviate, Pinecone, Qdrant, FAISS, ChromaDB ▸ Models: OpenAI, Claude, Gemini, fine-tuned open-source ▸ Automation: n8n, Make (Integromat), Zapier, Supabase, Twilio ▸ Backend: Python (FastAPI, Flask), TypeScript/Node (NestJS, NextJS), PostgreSQL, MongoDB ▸ MLOps & Cloud: Docker, Kubernetes, KEDA, CI/CD, Airflow, MLflow; AWS (SageMaker, Lambda), Azure ML / Azure OpenAI, GCP, serverless GPU ▸ CV & Data: OCR optimization, vision-language models, Stable Diffusion, web scraping ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ WHY CLIENTS PICK ME ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ▸ Ships to production. I build AND deploy. You get systems that run 24/7 and scale, not a prototype someone else has to finish. ▸ Proven track record. Top-Rated Plus, 100% Job Success, enterprise and healthcare AI delivered end-to-end. ▸ Innovation-driven. I bring the latest (MCP, adaptive RAG, new model releases) into production. ▸ Cost-conscious. High-performance AI that optimizes spend without compromising quality. ▸ Quality-first. Production-grade code, proper testing, and evaluation built in. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Building an AI product, or need one taken from prototype to production and scaled reliably? Send me the brief and I'll tell you exactly how I'd approach it.

I'm an LLM and ML engineer with about 3 years of experience building production AI systems. Most of my work is on agents, RAG pipelines, fine-tuned open models, and the evaluation setup that tells you whether any of it is actually getting better. A pattern I see in projects that come to me: the output looks fine on a few examples, breaks on others, and there's no way to measure when or why. Setting up that measurement loop is usually where I add the most value, and it's where a lot of LLM and agent projects get stuck. What I'm working on right now: - ML engineering at Manana Labs, building agentic systems and LLM applications for client work. - An agentic orchestration system for a Stanford research lab. Replaced an unreliable prior implementation with a Claude Agent SDK pipeline and a custom eval harness, delivered in about 3 months. - Maintaining unlearn-diff, a small open-source PyPI library for machine unlearning in diffusion models. Areas I'm strongest in: - Agentic systems with LangGraph or Claude Agent SDK - RAG: chunking, hybrid search, reranking, grounding evaluation - Eval harnesses: LLM-as-judge, deterministic checks, regression on golden sets - Fine-tuning open models (Gemma, Llama, Qwen) with LoRA or QLoRA, on Runpod or AWS GPUs - Production deployment on AWS (Lambda, ECS, EC2, Cognito, CloudFormation), Docker, CI/CD, monitoring Past results worth mentioning: - Self-hosted invoice extraction with OCR and Gemma-27B running on g6e.xlarge. Took over a single-prompt pipeline that wasn't working, broke the task into sub-steps, and tuned prompts against an eval harness. Accuracy crossed 90%. - Stanford research lab tooling. Orchestrated complex scientific analysis tools through agents and iterated against an eval suite I built from scratch. - Document Information Extraction at MLExperts. Reached 90%+ benchmark accuracy on production traffic. - Law Baje. Domain-grounded legal RAG for Nepali law, around 80% grounding accuracy. - AI Crusade 2023, Environment track winner (transformer predictive maintenance). 1st place at SXC Sandbox Hackathon 2024. A recent client review on Upwork: "I really liked working with Bala Ram. He delivered all his tasks before deadline and what I really liked about him is his coding structure and communication." Stack I work with regularly: Python, PyTorch, LangGraph, Claude Agent SDK, Transformers, OpenCV, FastAPI, Django, Postgres, Redis, Celery, AWS (Lambda, ECS, EC2, DynamoDB, Cognito), Docker, GitHub Actions, Weights & Biases, Langfuse, vLLM, Runpod. How I work: - Async-first. I send proactive updates so you're not chasing me for status. - I back claims about model performance with eval numbers, not vibes. - I take ownership end to end: design, build, evaluation, deployment, monitoring. - I'm available 20-30 hours a week and can overlap US business hours. Best fit for: - Founders and teams shipping AI or ML products who need a senior engineer to make an agent, model, or RAG pipeline reliable. - Rescue projects where a previous attempt didn't perform. - Eval and observability buildouts on existing systems. - End-to-end agentic feature work. Probably not the right fit if: - The work is mostly UI/UX or design. - There's no clear definition of what "working" means and you don't want to define one. If you have a problem in this area, send the spec or a Loom. I'll come back within 12 hours with a concrete plan or an honest "not the best fit for me".

🔹 AI-Powered Solutions | Scalable Deployments | API Integrations | Data Engineering With 2+ years of experience in LLM deployments, 4+ years of experience in Generative AI solutions, 9+ years in API development, and data engineering, I specialize in building, optimizing, and deploying AI-driven applications that scale seamlessly. From high-performance APIs to AI-powered automation and cloud-native solutions, I deliver production-ready implementations that are efficient, secure, and scalable. 💡 How I Can Help You: ◉ PRODUCT DEVELOPMENT & API DESIGN ✅ FastAPI & Flask – High-performance Async API development ✅ Django – Secure, scalable web applications ✅ Google Analytics & BigQuery – Data tracking & insights ◉ LLM & GENERATIVE AI SOLUTIONS ✅ LLM Deployment – GPT, Retrieval Augmented Generation (RAG), Spark NLP ✅ AI Integrations – Text generation, Document AI, Entity Extraction ✅ Vector Databases – ChromaDB, ElasticSearch, MongoDB ◉ CLOUD Services & DEPLOYMENT ✅ AWS Bedrock and Sagemaker ✅ Docker & Linux – Containerized microservices, multi-stage builds ✅ Kubernetes & Orchestration – Scalable, resilient AI/ML workloads ✅ AWS Services & GCP – EC2, S3, ECS, Lambda, Datadog monitoring ◉ DOMAIN-SPECIFIC EXPERTISE ✅ US Healthcare Solutions – AI-driven analytics & automation 🔍 Looking for Long-Term Projects (~30 hrs/week) in My Expertise 🙋🏽 About Me I’m a full-time freelance developer from Nepal ⛰ with a passion for building scalable, maintainable, and production-ready solutions. Whether it's LLM deployments, cloud infrastructure, or API development, I bring efficiency and expertise to every project. Let’s create something amazing together! Message me to discuss your project. 🚀

🚀 Passionate and Professional Data Scientist & AI Engineer I specialize in Generative AI, Data Science, and API Development, with proven expertise across diverse industries, including non-profits, private enterprises (product/service), and medicine. Harnessing a deep understanding of how Generative AI is transforming businesses, I design innovative and impactful solutions—whether it's creating intelligent chatbots, building advanced assistants for precise and actionable outcomes, or automating workflows to eliminate manual, non-productive tasks and boost efficiency. 🌟 Expertise I specialize in leveraging advanced machine learning techniques to solve real-world problems, with a strong focus on: Computer Vision: (e.g., U-Net, Segment Anything) Natural Language Processing: Large Language Models (LLM), BERT, Deep Neural Networks, HuggingFace Generative AI: LLM, Agentic AI, Multimodal models, RAG API Development: End-to-end development and integration in business processes 💻 Technical Proficiency Platforms & Tools: AWS SageMaker, LangGraph Platform (Graph Deployment), Docker, CI/CD Frameworks: PyTorch, TensorFlow, LangChain, LangGraph Database Management: MongoDB, VectorDB, GraphDB Collaboration: GitHub, agile workflows Monitoring and Observability: LangSmith