All public repositories on my GitHub, ordered from strongest to lightest: production-style, deployed work first.
01Multi-agent LLM · Full-stack
FinSight AI: Multi-Agent Equity Research Assistant
An AI research analyst for Indian (NSE) stocks. Ask about a company, a comparison or your portfolio and get a cited answer in which every claim links back to the tool call or filing page it came from. The BUY/HOLD/SELL verdict comes from a deterministic scorecard, not the LLM.
- LangGraph pipeline: an orchestrator routes each question to one of 7 intents, research, sentiment and risk ReAct agents run in parallel, and synthesis returns schema-validated answers (Pydantic, strict JSON schema with a function-calling fallback). A validator checks citations and triggers one revision when needed.
- Rule-based 9-factor scorecard (trend, momentum, RSI, valuation, growth, FinBERT news sentiment, Prophet and XGBoost signals). The LLM explains the verdict but can't change it, which fixed a model that answered HOLD to about half of all questions.
- Hybrid RAG over 15 annual reports and earnings calls: 2,563 pages turned into 7,087 table-aware chunks, BM25 + bge embeddings with reciprocal-rank fusion. Recall@5 0.74 and MRR 0.53 at about 100 ms p50 on a 34-question, page-verified benchmark.
- Evaluation harness with a 25-case eval set, LLM-as-Judge, citation and number-grounding metrics, and latency tracking. A NIFTY 50 backtest showed the price-only signals had no edge (only RSI helped), and I reported that rather than tuning on the test set.
- Human-in-the-loop review: analysts correct answers in plain English, wrong corrections get pushback with evidence, accepted ones are versioned and become cited context in later answers, and approved answers publish as research notes.
- React 19 + TypeScript frontend with a live agent trace and Sources panel. FastAPI streams over SSE, with Celery/Redis jobs and PostgreSQL. The whole stack is deployed on AWS EC2 with Docker Compose and Caddy HTTPS, and works with any provider (Groq, OpenRouter, Gemini, Anthropic).
LangGraphRAGFastAPIReactTypeScriptPostgreSQLChromaDBCeleryDockerAWS
02Full-stack · LLM
Socratic LLM Tutor
An AI tutoring platform that builds critical thinking through adaptive questioning, misconception diagnosis and progressive scaffolding instead of handing out answers.
- Python REST API for student management, persistent chat sessions and response streaming, with dynamic model routing via the OpenAI SDK on OpenRouter.
- React 19 + TypeScript single-page frontend with Tailwind CSS and react-router-dom.
- JWT admin authentication (OAuth2 bearer flow, bcrypt hashing) and an analytics dashboard with searchable logs, full transcript views and one-click export.
- PostgreSQL storage; containerised with Docker Compose.
ReactTypeScriptFastAPIPostgreSQLOpenRouterDocker
03NLP · Search
IITGN E-Journal Semantic Search
Natural-language discovery for the IITGN e-journals portal. Ask for “AI helping doctors” and get relevant journals, not keyword misses. Built entirely on free, open-source tools.
- Encoded 10K+ academic journals into 384-dimensional embeddings with Sentence Transformers (all-MiniLM-L6-v2).
- FAISS IndexFlatIP similarity layer for cosine nearest-neighbour retrieval, combined with SQL metadata filtering in a hybrid ranking.
- CI/CD pipeline with automated Bandit, Safety and TruffleHog security scans.
- Shipped as a Dockerised FastAPI microservice with an embeddable JavaScript widget, hosted on Railway.
PythonFastAPIFAISSSentence TransformersSQLDockerCI/CD
04Full-stack · In development
IITGN Alumni & Donor CRM
An open-source, modular CRM for IIT Gandhinagar's Alumni Relations and Institute Advancement offices, designed to replace fragmented spreadsheets.
- Planned modules: auditable people registry, donation ledger, engagement workspace, reporting and an alumni self-service portal.
- FastAPI modular-monolith backend with React + TypeScript internal frontend.
- Engineering-first process: architecture docs, ADRs, release checkpoints with acceptance criteria and runbooks.
- Actively developed; currently at the M0 foundation milestone.
FastAPIReactTypeScriptSystem design
05Agentic AI
Agentic Email Manager
An autonomous Gmail assistant that reads your inbox, classifies email, extracts tasks and drafts replies, all behind a Streamlit dashboard.
- Gemini-powered classification into Action Required, Informational and Spam, plus task, deadline and priority extraction.
- Context-aware reply drafting with tone detection and Google Calendar event creation from meeting mentions.
- Priority engine and decision log with reasoning and confidence scores; every action needs manual approval.
- Gmail and Calendar APIs, Google OAuth 2.0, SQLite via SQLAlchemy, background monitoring loop.
PythonFastAPIGeminiGmail APIStreamlitOAuth 2.0
06Computer vision
Satellite Land-Use Classification
Land-cover classification of Delhi NCR from Sentinel-2 RGB imagery using a ResNet18 CNN.
- End-to-end pipeline: GeoJSON-based spatial filtering, patch extraction, automated label construction and train/test split.
- Derived dominant land-cover labels from raster data to build a supervised learning dataset.
- Trained and evaluated ResNet18 with accuracy, F1-score and confusion-matrix analysis.
PyTorchResNet18Sentinel-2GeoJSONscikit-learn
07Machine learning
Real Estate Rent Estimator
A Streamlit app that estimates rental prices from property attributes, geolocation and market features.
- Data preprocessing with categorical encoding and feature engineering, plus visual analysis across categories.
- Random Forest Regressor tuned with Grid Search, Random Search and Bayesian Optimization (Optuna); compared via cross-validation and experiment tracking.
- Dockerised and deployed on Hugging Face Spaces.
PythonRandom ForestOptunaStreamlitDocker
08Deep learning
Anomaly Detection on Tennessee Eastman Process
An LSTM autoencoder that learns normal industrial-process behaviour and flags faults from reconstruction error.
- Stacked LSTM encoder-decoder in PyTorch over 50-timestep sliding windows of 52 process variables and 21 fault types.
- Threshold set at mean + 3σ of validation reconstruction error; evaluated with ROC-AUC, precision/recall/F1 and detection delay.
- Modular, reproducible pipeline: one command runs data loading, training, evaluation and six visualisation types.
PyTorchLSTMAutoencoderTime series
09Biomedical signals
Sleep Apnea Detection from Respiration Signals
A deep-learning pipeline that classifies breathing anomalies such as hypopneas and apneas from overnight sleep recordings.
- Processed 8-hour, multi-modal recordings (nasal airflow, thoracic movement, SpO2) from 5 participants with Butterworth band-pass filtering and interpolation.
- Built 30-second windowed datasets with 50% overlap and a custom 1D-CNN (SleepCNN) in PyTorch.
- Evaluated with Leave-One-Participant-Out cross-validation; automated signal visualisation.
PyTorch1D CNNSignal processingNumPy
10Data analytics
Product Hunt Trends Analyzer
A Streamlit dashboard that explores Product Hunt market dynamics to inform product-launch strategy.
- Pulls trending products ranked by upvotes from the Product Hunt API.
- Animated gradient visualisation of the most-used tags, with chart export and a mobile-responsive UI.
PythonStreamlitREST API
11EdTech · LLM
Personalized Micro-Learning Engine
An adaptive Python-learning platform that keeps lessons short and picks what to learn next based on learner performance.
- Gemini 2.5 Flash/Pro assess competence and generate a dynamic learning path with instant quiz feedback.
- FastAPI backend with HTML/CSS/JS frontend, deployed on Vercel; includes a freemium business model.
PythonFastAPIGeminiVercel
12Web
This Portfolio Website
The site you are reading: hand-written HTML, CSS and JavaScript, responsive with light and dark themes, hosted on GitHub Pages.
- No frameworks or build step; accessible, mobile-friendly and fast.
HTMLCSSJavaScriptGitHub Pages