# Dhruv Mishra — Projects

Source: https://whoisdhruv.com/projects

Cropio is my privacy-conscious AI portrait cropper that turns raw photos into polished headshots. Live at cropio.whoisdhruv.com. Architecture: Next.js + TypeScript frontend with a Python FastAPI backend. Uses YOLO11 pose estimation plus face-orientation detection to capture different angles of the face and generate multiple headshot crop suggestions, with deterministic geometry fallbacks, an interactive drag-resize editor with aspect-ratio presets, and full-resolution browser exports. Image descriptions are generated via NVIDIA multimodal LLM APIs and indexed locally in vector IndexedDB for semantic session search over saved crops — no server-side image storage, everything stays in the browser. Supports full local session management. Stack: Next.js, TypeScript, FastAPI, YOLO11 Pose, IndexedDB, NVIDIA APIs, OpenCV, Python. Year: 2026, ongoing. Repo: github.com/Dhruv-Mishra/Cropio-ImageEditor.

Fluent UI Android is the official Microsoft native Android library enabling developers to build uniform Microsoft 365 experiences. I was its primary owner — driving 10 major releases used by 10+ partner teams including Copilot, Office, Outlook, OneDrive, and Teams, and I migrated and owned the publishing pipelines for Fluent UI System Icons. The library is Kotlin/Java with Jetpack Compose, XML layouts, Gradle/Maven builds. It ships official Fluent design tokens, typography styles, and custom controls for seamless integration with the Microsoft ecosystem. Reaches 100M+ end users across M365 apps. Stack: Kotlin, Java, Android SDK, Design Systems, Clean Architecture, API Design. Repo: github.com/microsoft/fluentui-android.

Jarvis is my voice-to-voice AI agent that picks up the phone, holds a full human-sounding conversation, and actually operates a website on the caller's behalf via tool calling. It's a personal side-project I built end to end on my own time — NOT a Microsoft product, not part of the Microsoft 365 shell, and not affiliated with any of my employer's work. It's pitched as an alternative to traditional customer-support and dispatch agents — call in, and it can navigate pages, fill forms, open maps, send quotes, schedule callbacks, look up records, negotiate, and hold a complete conversation end to end. Live demo at jarvis.whoisdhruv.com (anyone can try it). Source code is public at github.com/Dhruv-Mishra/AudioControlledAgenticWebsite. Architecture: vanilla HTML/CSS/JS frontend with an AudioWorklet capturing 16 kHz PCM, a Node.js 20+ backend that bridges a long-lived WebSocket to a live AI agent, and tool schemas that let the model drive real DOM actions. UX is built around a phone-call metaphor — Place Call / End Call, ambient call-center noise, 300–3400 Hz phone-line band-pass compression, persona switching (Professional, Cheerful, Frustrated, Tired, Excited), barge-in support, and native VAD. Seamless SPA navigation: the WebSocket, AudioContext, and microphone live in a single long-lived shell so the call doesn't drop when the user moves between pages. Token-efficient — no audio is uploaded until the call starts, sliding-window compression at 80k tokens, and a stable system-prompt + tool-schemas prefix for prompt caching. Stack: JavaScript, Node.js, WebSockets, AudioWorklet, Tool Calling. Year: 2026, ongoing. Repo: github.com/Dhruv-Mishra/AudioControlledAgenticWebsite. Live demo: jarvis.whoisdhruv.com. Following the live-demo link from the projects page also unlocks the hidden "Phoned a Friend" sticker.

The portfolio website itself — a high-performance server-rendered site built with Next.js 16, React 19, TypeScript, Tailwind v4, and Framer Motion, with a hand-drawn sketchbook aesthetic (custom pencil/chalk cursor, paper textures, torn tape strips). Features an AI-powered chat, an interactive retro terminal, a guestbook, an unlockable sticker system, and a custom command palette. Infrastructure: georedundant deployment across Oracle Cloud, GCP, and Azure VMs with custom traffic routing, Nginx reverse proxies, Cloudflare caching, and separate GitHub Actions CI/CD pipelines. Everything free except the domain. Stack: Next.js, TypeScript, TailwindCSS, Framer Motion, Azure, GCP, Oracle Cloud, Nginx. Year: 2025, ongoing. Repo: github.com/Dhruv-Mishra/portfolioWebsite.

Bloom Filter Research — my undergraduate research at IIIT Delhi's DCLL lab under Prof. Bapi Chatterjee, focusing on optimizing Counting Bloom Filters for high-concurrency systems. Achieved a 300% throughput increase via relaxed synchronization techniques in C++, benchmarked against state-of-the-art concurrent filter implementations with minimal impact on false positive/negative rates. Published in the IIIT Delhi repository. Stack: C++, Bloom Filters, Concurrency, Optimization, Data Structures. Year: 2024, 8 months. Link: repository.iiitd.edu.in/jspui/handle/123456789/1613.

Course Similarity Evaluator — an intelligent Python tool designed to detect redundant course content across university curriculums. Built a fuzzy matching pipeline with text similarity algorithms over course syllabi, identifying redundant modules with configurable similarity thresholds so students can avoid retaking equivalent coursework. Originally an engineering project at IIIT Delhi under Prof. Dhruv Kumar. Stack: Python, Fuzzy Logic, NLP, Data Analysis, Algorithm Design, scikit-learn. Year: 2023, 2 months. Repo: github.com/Dhruv-Mishra/Course-Similarity-Evaluator.

Instant Vital Checkup (IVC) — a contactless, computer-vision-powered health screening kiosk that automates patient triage. Using OpenCV and MediaPipe in a real-time pipeline, it calculates height, weight, BMI, and pulse from a single camera at a distance, drastically reducing wait times. Stack: Python, OpenCV, MediaPipe, Computer Vision, HealthTech, Real-time Processing. Year: 2023, 4 months. Repo: github.com/Dhruv-Mishra/Instant-Vital-Checkup-IVC.

AtomVault — a secure, ACID-compliant banking database built for high-reliability transactions. Features full transaction rollback and recovery, multi-user architecture, and strict role-based access control with admin, teller, and customer roles. Stack: Java, MySQL, JDBC, OOP, ACID Compliance. Year: 2022, 2 months. Repo: github.com/Dhruv-Mishra/AtomVault.

Hybrid Entertainment Recommender — an age-aware, context-sensitive movie recommendation engine built for family movie nights. Combines collaborative filtering and content-based filtering into a hybrid engine, with age-appropriateness scoring for family-safe recommendations and a group preference balancing algorithm for multi-user sessions. Stack: Python, scikit-learn, Collaborative Filtering, ML System Design. Year: 2023, 3 months. Repo: github.com/Dhruv-Mishra/Age-and-Context-Sensitive-Hybrid-Entertaintment-Recommender-System.