Hi, I'm @subash
Building production-ready AI systems for real business outcomes.
Remote AI/ML engineer focused on GenAI products, RAG systems, and LLM reliability.
Work Experience
End-to-end ownership across research → implementation → deployment → evaluation, collaborating with cross-functional teams to translate business goals into deliverable AI features.
Scopic Software LLC
Remote AI/ML Engineer
Oct 2024 - Present
- •Contributed to development and maintenance of a conversational AI platform with AI-driven features
- •Led development of GenAI chatbot with LangGraph, RAG pipelines, and Qdrant for intent detection and contextual responses
- •Conducted R&D on audio feature extraction using Parselmouth, Librosa, and Praat
- •Estimated AI project scope with man-hour granularity for internal and client-facing projects
- •Maintained AI service layer of an assessment platform: question generation, grading logic, prompt engineering
- •Implemented evaluation pipelines for chatbot and RAG performance assessment
- •Contributed to keyword research and SEO SaaS platform as full-stack developer
Tech Stack
GenAI Chatbot
Led development and deployment of a GenAI chatbot integrated with CRM for intent detection, user info extraction, contextual responses, and analytics tracking.
- Conversation classification and routing via LLM
- Intent detection and user info extraction
- CRM integration for sales pipeline
- SEO tracking: referrer, session time, last visited page
- RAG pipeline for contextual responses
- Evaluation pipeline for response quality
Conversational AI Platform
Implementing and maintaining AI-driven features for a conversational AI platform, resolving production bugs and ensuring system reliability.
- AI-driven conversational features in production
- Production bug resolution and system reliability
- Conversational quality across deployments
- End-to-end feature implementation
- Cross-functional collaboration with remote teams
Assessment Platform AI Services
Maintained and expanded the AI service layer of an assessment platform: question generation, grading logic, prompt engineering, and end-to-end GenAI integration.
- Various types of question generation
- Automated grading logic
- Prompt engineering optimization
- End-to-end GenAI integration
Peace Nepal Dot Com
AI/ML Engineer
Jul 2024 - Nov 2024
Designed and developed chatbots for Banking, Travel, and Customer Support use cases. Implemented RAG and multi-agent frameworks for enhanced responses. Integrated databases and knowledge sources. Deployed interactive web chat applications.
Icebrkr AI Solutions
AI/ML Engineer
Mar 2024 - Jun 2024
Engineered real-world AI solutions using machine learning. Built scalable ML models for data analysis and automation. Optimised and deployed models for production in cloud environments. Collaborated with cross-functional teams.
Contentio Lab Pvt. Ltd.
Kathmandu, NepalData Analyst
Aug 2021 - Apr 2022
Analysed large datasets using Python to identify trends and patterns. Created data visualisations using Matplotlib and Seaborn. Developed data pipelines leading to 20% reduction in processing time. Supported data-driven decisions across marketing and sales via customer segmentation and churn prediction.
iMark Private Limited
Frontend Intern
Oct 2020 - Jan 2021
Developed a to-do web application using React.js with task creation, management, and completion features. Implemented a CRUD backend for data persistence. Integrated secure login component using React for user authentication.
Budhanilkantha Education Services
Kathmandu, NepalTutor
Oct 2018 - Apr 2019
Tutored A-Level Computer Science. Helped failing students obtain passing grades. Developed personalised lesson plans and provided targeted exam preparation strategies.
Projects
notsubashActivity Recognition
Classified human physical activities (walking, jogging, sitting, typing, etc.) using XGBoost on accelerometer and gyroscope data from smartphones and watches. Hyperparameter tuning via random search achieved >85% accuracy.
Steam Video Game ML
Explored the evolution of game genres on Steam using network analysis. Identified emergent sub-genres, influential genre nodes, and key drivers of player engagement including ratings, achievements, and pricing models.
Wikipedia Adminship
Investigated social network dynamics within Wikipedia's administrator election process. Calculated centrality measures, performed community detection via hierarchical clustering, and analyzed voting blocs.
Stock Variables Analysis
Exploratory data analysis of stock prices across exchanges. Identified significant variables affecting closing prices through correlation analysis, regression, feature selection, and data clustering.
Floki eCommerce
Full-stack e-commerce application built with the MERN stack. Features include product browsing, searching, filtering, cart management, user authentication, order processing, and payment integration with Redux state management.
AI/ML Toolkit
Production ExperienceMy daily work involves building and maintaining AI-driven features in production, from conversational AI systems to evaluation pipelines that ensure quality and reliability.
- ›RAG pipelines with LangGraph and Qdrant for intent detection, contextual retrieval, and user info extraction
- ›LLM evaluation pipelines assessing chatbot and RAG performance: hallucination rate, accuracy, cost, latency
- ›Audio feature extraction using Parselmouth, Librosa, and Praat for voice-based AI applications
- ›Full-stack AI services: question generation, grading logic, prompt engineering, and end-to-end GenAI integration
Sharing
Publications
Wikipedia Adminship Network Analysis
Investigated social network dynamics within Wikipedia's administrator election process using centrality measures and community detection.
Recommendations
LinkedIn Posts
Blog
Technical Notes
Bite-sized things I learned building real systems.
Hybrid Retrieval in RAG
Run a metadata-filtered search (e.g. content_type + technology) and an unfiltered similarity search in parallel. Merge filtered-first, dedup by chunk ID. Filtered results give precision; unfiltered results fill context gaps the filters miss.
Query Rewriting Before Vector Search
Rewrite conversational messages into standalone search queries using an LLM + chat history before hitting the vector store. "Yeah what about that?" becomes "What mobile development services are available?" Single biggest retrieval quality improvement I've made.
HPSS for Harmonic-to-Noise Ratio
Instead of autocorrelation-based HNR, split the STFT into harmonic and percussive components via librosa's HPSS (median filtering on the spectrogram), then compute 10·log₁₀(E_harmonic / E_percussive). Different from Praat's HNR but captures voice quality well.
Education
Master of Science in Data Science
Statistics, Machine Learning, Social Network Analysis, Data Science Applications. Thesis: Machine Learning the Steam Video Game Database. Grade: Merit.
BSc (Hons) Computing
Computer Science Fundamentals, Software Engineering, Databases, Networking, Cloud Computing. Thesis: Floki, An eCommerce Web Application. Grade: 2:1.
