- 🎓 Computer Science student at JIIT Noida (2024–28)
- 💻 I love building intelligent systems that interact, automate, and improve lives
- 🇮🇳 Mission-driven to use tech for India's advancement in defense and wellness
- 🧪 Constantly experimenting with AI, ML, automation, and hardware integration
Python, YOLOv8, DeepFace (Facenet512), FAISS, MediaPipe, SQLite, SMTP, SciPy
Project Drishti, an end-to-end AI surveillance pipeline that detects littering in real-time using computer vision, identifies violators through facial recognition, and automatically dispatches digital e-challans to enforce civic accountability under the Swachh Bharat mission.
- Engineered a temporal state machine that tracks person-garbage relationships over time (UNTRACKED → ATTACHED → DETACHING → MONITORING → LITTERING_CONFIRMED) to accurately infer littering intent and drastically reduce false positives in crowded scenes.
- Integrated YOLOv8 for real-time multi-object detection (humans, garbage, dustbins), custom centroid-based tracker for temporal tracking, and MediaPipe + DeepFace for face isolation and identity embedding generation.
- Built a high-speed facial recognition pipeline combining FAISS vector indexing (Facenet512 embeddings) with SQLite lookups to match violators against registered identities, enabling instant violation attribution.
- Implemented automated e-challan generation via SMTP, complete with violation evidence (timestamp, image, violator details, Rs. 500 fine notice) supporting Swachh Bharat (Clean India) initiatives at scale.
Python, PyTorch, DenseNet-121, Grad-CAM, FastAPI, FAISS, Groq Api
Dr. Chakshu, an end-to-end AI-powered clinical decision support system designed to assist clinicians and patients in chest X-ray analysis by combining Computer Vision, Explainable AI, Retrieval-Augmented Generation, and structured report generation.
- Built a multi-stage AI diagnostic pipeline for chest X-rays using a fine-tuned DenseNet-121 backbone trained on 57K+ NIH CXR-14 images, achieving a mean AUC-ROC of 0.7735 across Cardiomegaly, Infiltration, and No Finding classes.
- Engineered an explainability layer via Grad-CAM exposed through a FastAPI endpoint that returns the highest-probability disease prediction alongside a Base64-encoded heatmap overlay, highlighting clinically relevant lung regions.
- Designed a RAG-augmented report generation module using FAISS as a local vector store with sentence-transformers embeddings and the Groq API as the LLM, grounding AI-drafted structured radiology-style findings and recommendations.
Python, Sarvam AI, Groq, Twilio, Silero VAD
Stock Market Voice, an AI-powered voice calling system for brokers that schedules and initiates automated calls to clients, delivering personalized daily stock purchase summaries with real-time market intelligence.
- Built an AI Voice Agent that makes automated concurrent outbound phone calls to clients with personalised stock portfolio summaries; implemented a LangGraph ReAct agent with 4 live market data tools (live price, metrics, history, market status) and zero-hallucination architecture by injecting per-client trade data from Supabase into the LLM system prompt at runtime.
- Developed an end-to-end real-time voice pipeline (Twilio Media Streams → Silero VAD → Sarvam Speech-to-Text → LLM → Sarvam Text-to-Speech → WebSocket) supporting multiple languages with streaming TTS playback that begins within the first audio chunk; engineered a barge-in interruption system detecting user speech in 300ms that instantly cancels in-flight audio and resets the conversation state.
- Implemented robust multi-turn conversational memory using LangGraph InMemorySaver with per-call thread isolation, and designed LangGraph state pruning via RemoveMessage to recover gracefully from mid-tool-execution interruptions, ensuring zero agent crashes across concurrent calls.
Python, OpenCV, DeepFace (Facenet512), FAISS, SQLite, Silero VAD, Groq LLaMA, Sarvam STT/TTS
Vayu, an interactive face recognition bot that recognizes faces via FAISS embeddings and runs a personalized, streaming voice conversation pipeline (STT → LLM → TTS).
- Developed a real-time AI booth assistant for a college fest (JIIT Converge'26) that identifies attendees via FAISS-based cosine-similarity search over Facenet512 face embeddings and delivers fully personalized welcome interactions in under 2 seconds.
- Built an end-to-end voice pipeline (Sarvam AI Speech-to-Text → LLaMA 3.3-70B → Sarvam AI Text-to-Speech over WebSocket) with streaming audio playback that begins speaking within ~0.4 seconds, creating fluid human-like dialogue.
- Engineered concurrent execution using Python's
ProcessPoolExecutorto run face recognition in an isolated process alongside live camera capture (OpenCV) and a Silero VAD audio thread, ensuring low-latency multimodal interaction stability.
Python, OpenCV, Silero VAD, PyAudio, Sarvam AI (STT/TTS), Groq Qwen (Vision)
JARVIS — Secret Key Extraction is a fun, high-energy AI game built for JIIT fresher induction interactions.
A fresher stands in front of the camera and talks to JARVIS, whose job is to guard a hidden secret key while roasting the player in real time.
Game Rules:
- At the start of every round, JARVIS picks a random secret key (e.g., Bunk, Maggi, Proxy, Attendance).
- The player must convince/trick JARVIS to reveal that key through conversation.
- JARVIS stays aggressive and refuses direct attempts, making prompt strategy the core challenge.
- If the key is extracted, JARVIS declares the player the winner and the session ends.
Technical Highlights:
- Built a real-time multimodal pipeline combining live webcam + voice interaction.
- Audio stack: Silero VAD → Sarvam STT → Groq Qwen Vision → Sarvam streaming TTS for low-latency, natural turn-taking.
- Vision support captures the best webcam frame during speech, enabling contextual and appearance-aware responses.
- Added language adaptation for English/Hindi/Hinglish and dynamic TTS voice behavior.
- Implemented gameplay-state logic for robust win detection and automatic shutdown on successful key extraction.
n8n, Twilio, Sarvam AI, Gemini, Python
Customer Recovery Automation Tool, an AI-powered customer recovery workflow that transforms vague negative reviews into actionable business insights through intelligent voice conversations.
- Built an intelligent review analysis pipeline that detects negative and unclear customer feedback, then automatically triggers personalized AI voice calls to gather detailed feedback through natural dialogue.
- Integrated a multi-stack voice system using Twilio for calling, Sarvam AI for STT & TTS, and Gemini AI for review analysis, orchestrated seamlessly through n8n workflows.
- Engineered an AI-driven escalation logic system where the workflow autonomously decides when to escalate for human intervention, transforming passive feedback analysis into proactive customer recovery operations.
Python, FastAPI, Scikit-learn, Linear Regression
Sleep Debt Predictor, a web-based ML system that estimates user sleep debt by analyzing facial fatigue indicators (eye redness, dark circles, yawn frequency) captured via a real-time browser video stream and converting them into personalized recovery recommendations.
- Trained a Scikit-learn Linear Regression model on a self-collected dataset, achieving R² 0.87, MAE 0.80 hours, and RMSE 1.02 on the test split.
- Built an end-to-end pipeline integrating a FastAPI backend with Google Gemini API for feature extraction, and implemented Web Speech API for AI-generated personalized voice feedback.
🕵🏻♂️ Spy AI
Flutter, FastAPI, Dart, Python, Gemini, PostgreSQL
Spy AI, an AI-powered full-stack mobile app using Flutter (Dart) frontend and FastAPI (Python) backend that turns your phone into a searchable, lifelong memory by recording and indexing your daily conversations.
- Implemented a background recorder with real-time audio capture, transcription pipeline, structured database storage and full-text search; integrated a Gemini 2.5 Flash-powered chatbot that answers memory queries using contextual retrieval.
- Delivered end-to-end production features including secure storage, privacy controls, meeting-level metadata, and cross-platform deployment—empowering users to recall details, verify claims, and act on past discussions instantly.
Flask, Html, Twilio, Gemini
AI Calling, a full-stack voice-first application with an HTML frontend and Flask backend that enables natural, real-time phone conversations with an AI assistant.
- Integrated telephony providers and built real-time speech-to-text and text-to-speech pipelines, plus automated call initiation to allow users to talk directly to an LLM-driven agent over voice.
- Implemented an emergency auto-call feature to send urgent voice messages when users cannot place calls manually, and delivered secure call handling and scalable backend services.
🍽️ Flavor Match (Live Demo)
FastAPI, Gemini, Html, CSS, JavaScript
Flavor Match, a full-stack family food recommendation web app using HTML/CSS and JavaScript on the frontend with a FastAPI + PostgreSQL backend; deployed a live demo on Vercel with responsive UX.
- Designed and implemented a normalized relational schema (Family → Member → Food) applying primary/foreign keys, UNIQUE/NOT NULL constraints and cascade operations, with timestamped logs.
- Implemented streamlined family/member registration and daily food-logging flows, plus an SQL-driven recommendation pipeline (rule-based / lightweight AI) to deliver personalized meal suggestions.
Languages
Frameworks & Tools
AI & APIs
Crafted with 💻 logic, ☕ passion, and 🇮🇳 heart