Melvin Joshua

A multilingual RAG chatbot answering real admission queries.

The official admission helper for Loyola College, Chennai, grounded on the current prospectus and built end to end by a two member team. It answers programme, fee, date, and eligibility questions in English, Tamil, and Hindi.

• Hybrid retrieval: keyword search and semantic vectors fused with Reciprocal Rank Fusion, plus a structured facts layer for authoritative dates, fees, and contacts
• Multi-provider LLM orchestration with streaming and status-aware failover; answer caching cuts repeat embedding and LLM cost
• Voice input, text to speech, in-place translation, and dynamic follow-up suggestions
• Defensive layer: per-IP rate limiting, prompt injection guards, sanitized error boundaries, and a token-gated admin dashboard

Point it at any repository, get back a container that runs.

An agentic tool that clones a repository and generates a working Dockerfile through a self healing build loop: the model reads its own build errors, revises, and retries until the container runs. Zero manual Dockerfile writing.

• Self healing loop: build, read the error, patch the Dockerfile, rebuild
• Reasons about the actual build failure, not just surface syntax
• Works across languages and frameworks it has never seen before
• Streamlit interface: paste a repo URL and watch it iterate live
• Stops only once the container builds and runs, no babysitting

See where your money actually goes, with an assistant that reads your spending.

A personal finance web app for tracking expenses and income, with a conversational assistant that summarizes where the money goes and answers questions in plain language. I built the pages, implemented security across the app and chat, and shipped the agentic chat.

• Google OAuth sign-up and sign-in
• Agentic finance chat grounded in your own transaction history
• Security hardening across the whole app and the assistant
• Tiered AI routing: light models for routine queries, stronger ones when it matters

An AI tutor that lives inside WhatsApp.

A tutor delivered over WhatsApp that teaches Python, machine learning, and deep learning, meeting a learner right where they already chat.

• Structured lesson flow that adapts to the learner's pace
• Quizzes and multiple modes: learn, practice, and revise
• Covers Python, machine learning, and deep learning, all inside WhatsApp
• Self-hosted on a free-tier cloud VM for a permanently zero-cost deployment

Basketball outcomes, predicted while the game is still on.

A mobile first, cross platform basketball outcome predictor. Real time inference and outcome classification over live game data streamed from a real time database.

A meme factory assembled entirely from free APIs.

An automated meme generator built on workflow automation: prompt to template to caption, chaining the Telegram Bot API, the Imgflip API, and LLMs. Template matching with fuzzy similarity scoring and caption generation that recovers gracefully from vague prompts.

Explainable clinical ML that shows its working, patient by patient.

A research grade, explainable machine learning pipeline on liver disease data. Deep exploratory analysis, Weight of Evidence and Information Value feature analysis, and a decision tree plus logistic regression hybrid that reaches 0.96 AUC and beats a random forest.

• Per patient SHAP waterfall explanations
• Over one hundred test driven tests, hyperparameter tuning, and continuous integration
• Deployed interface for interactive exploration

1.2 million sensor rows, clustered, monitored, and served.

An industry grade rewrite of a distributed clustering pipeline over 1.2 million rows of accelerometer, gyroscope, and rotation data.

• Multiple clustering algorithms compared across 1.2M sensor rows
• Anomaly detection over accelerometer, gyroscope, and rotation streams
• MLflow experiment tracking, an API backend, and a web frontend
• Deployed end to end, from pipeline to interactive dashboard