Portfolio
Projects
All core logic built from scratch. No high-level framework shortcuts — each project reflects a problem I wanted to genuinely understand.
HTML5 / CSS3
Vanilla JS (ES6)
Web Workers
PDF.js
jsPDF
heic2any
- Description: A high-performance, privacy-first, 100% client-side file compression web application designed to run entirely in the browser. Compress JPEGs, PNGs, WEBP, HEIC images, and PDFs to target limits (e.g. 50KB, 100KB) quickly and safely.
- Hardware-Aware Concurrency Control: Automatically detects and leverages multiple CPU cores using Web Workers for ultra-fast multi-file batch processing. Includes a dynamic, load-aware concurrency control slider.
- Asynchronous Task Queue: Buffers large file batches to prevent browser memory spikes and tab crashes.
- Multi-Format Support: Handles JPEG, PNG, WEBP, Apple HEIC, and multi-page PDF documents.
- Smart PDF Downscaling: Decouples physical layouts from rendering resolutions to compress PDFs without changing physical page boundaries. Paints solid white backgrounds to remove transparency blackout artifacts in JPEGs.
- Two-Pass Compression for PDFs: Squeezes PDFs to a high-quality readable limit (minimum 500px and 0.5 quality) in the first pass, and enables a "Force to Target" option to override readability and forcefully compress the file to the strict target limit.
- Zero Server-Side Handling: All conversions and compressions happen inside the browser client. Your data never leaves your device.
Core Security
Custom Cryptographic Engine & Firmware Security
Completed
C / Python
AES / DES / RSA
SHA-256
IoT Security
- Description: Engineered custom implementations of AES, DES, and RSA from scratch in C and Python, demonstrating a deep understanding of core cryptographic primitives and bit-level data manipulation.
- Developed secure cryptographic hashing utilities, including SHA-256 and digital signature verification protocols, strictly utilizing standard libraries to ensure mathematical accuracy.
- Designed firmware architectures resilient to physical vulnerabilities, implementing defensive strategies against side-channel attacks and fault injection for embedded microcontrollers.
- Secured real-time data transmission for hardware modules by integrating custom lightweight encryption and device authentication protocols.
Virtual Experience · Mar 2026
Completed
Apache Kafka
H2 DB
Java
Real-Time Data
- Description: Engineered a real-time data visualizer using Apache Kafka and H2 DB during a virtual engineering simulation by Forage.
- Achieved sub-10ms response latency for high-frequency live data ingestion flows.
Hackathon Winner
TourAI: AI-Powered Smart Travel Planning Platform
[GitHub]
🥇 1st Place
FastAPI
React 18
Vite
Meta Prophet
Random Forest
Claude Sonnet 4
RAG
MySQL
Docker Compose
Nginx
Framer Motion
- Description: Full-stack AI travel planner that forecasts inflation-adjusted future trip costs, checks financial feasibility, generates EMI savings plans, and produces Claude-generated day-by-day itineraries.
- 6-module ML pipeline: Prophet → Random Forest → Installment Engine → Claude LLM → RAG → Alternative Finder. Prophet model trained on hand-curated 15-year Mizoram dataset.
- Random Forest classifier: 100 estimators, features — savings_burden_ratio, income_surplus, months_available, cost_ratio.
- JWT auth, user dashboard, gamified milestone savings rewards (₹250 bonus per 25% milestone).
- Full Docker Compose stack: MySQL + FastAPI + Nginx-served React. Team: Dilip Sahu (Full-Stack + Backend).
IoT & Firmware
NITMZ Bus Tracker: IoT Real-Time Campus Bus Tracking
[GitHub]
Active
ESP32
C++
NEO-6M GPS
SIM800L GSM
Python Flask
MySQL
Leaflet.js
- Description: End-to-end IoT system for real-time campus bus tracking at NIT Mizoram; transmits GPS telemetry over dual-network (GSM primary, Wi-Fi secondary) to a cloud Flask backend, rendered live on Leaflet.js map.
- 1-second telemetry cycle; ~50ms Wi-Fi latency, 300–800ms GSM latency; ~2.5m GPS accuracy.
- Dual-network failover with 60s GSM timeout; 10-packet circular buffer on ESP32 for outage continuity.
- JSON telemetry: lat/lng, speed_kmh, satellites, hdop, fix status, network type, ISO-8601 UTC timestamp. Role-based access: students, caretakers, admins.
- Hardware: ESP32 (240MHz), NEO-6M, SIM800L quad-band. Dilip's role: ESP32 Lead — all firmware (C++), hardware wiring, GPS integration.
Core CS
Java Digit Recognizer: Neural Network from Scratch
[GitHub]
Completed
Core Java (JDK 8+)
Java Swing
MNIST dataset
MVC pattern
- Description: Professional desktop app implementing a 3-layer feedforward neural network in pure Java — zero ML frameworks — trained on MNIST with an interactive 28×28 drawing canvas.
- Architecture: 784 input → 128 hidden (Sigmoid) → 10 output (Softmax); Xavier weight initialization. Manual backpropagation with Java primitive arrays; learning rate 0.05.
- 96% accuracy on 500 MNIST test images after 10,000-sample training.
- Pixel binarization: normalize /255 → threshold 0.3 → binary 0/1 (independently derived 1-bit quantization concept).
- Persistent weights via brain.dat serialization; thread-safe training via SwingWorker.
2026 — Present
CyberGuard-AI: On-Device AI Scam Call Detector
[GitHub]
In Progress
Python
Whisper.cpp
LoRA
GGUF
llama.cpp
Q4 quantization
- Description: Privacy-first, fully on-device scam call detector for India. No server, no cloud. Targets low-end Android phones and laptops (~480MB footprint).
- Pipeline: live mic → 16-parameter audio feature extraction → Whisper.cpp STT (Hindi/English/Hinglish) → LoRA adapter → phrase fast-path (200ms) → conversation arc tracker → Q4 GGUF LLM → SAFE/SUSPICIOUS/SCAM + confidence.
- Dataset: 1,398 samples via 7B LLM on Kaggle T4 + Indian High Court records + YouTube scam call recordings.
Next.js
TypeScript
Tailwind CSS
Django REST
MySQL
ChromaDB
RAG
LM Studio
- Description: Full-stack RAG ecosystem that turns unstructured book data into searchable, AI-powered knowledge with a privacy-first local LLM.
- Automated Selenium scraper ingests and normalizes book data.
- Dual-database: MySQL (relational metadata) + ChromaDB (vector embeddings).
- Privacy-centric: local LLM via LM Studio — zero cloud AI calls. Features semantic search, automated summaries, genre classification.
Python Flask
SQLAlchemy
MySQL 8.0
Flask-JWT-Extended
Jinja2
- Description: Web-based management system for nursing homes — resident lifecycle, room management, prescriptions, activity logging, family visits, three-tier RBAC.
- 3NF-compliant 15-table relational schema.
- Full resident lifecycle: admit → transfer → discharge → re-admit. Room types: private, joint, ward with live capacity tracking.
- Three-tier RBAC (admin, nurse, family) via JWT; stateless audit logging.
Data Science
Hyperliquid × Bitcoin Sentiment Analysis
[GitHub]
Completed
Python
Pandas
NumPy
SciPy
scikit-learn
Statsmodels
Plotly
- Description: Quantitative analysis of ~211K Hyperliquid perpetual futures trades merged with the Bitcoin Fear & Greed Index to study how sentiment regimes affect trader behavior.
- Statistical tests: Kruskal-Wallis, Mann-Whitney, Spearman/Pearson correlations, VaR/CVaR.
- 17 auto-generated dark-theme charts + auto-rendered Markdown report. Built for Primetrade.ai Data Science Hiring Assignment.
Desktop App
Water Sort Studio Pro
Completed
Java
Swing
DFS/BFS
Pruning
Memoization
- Description: Desktop puzzle solver for Water Sort puzzles using state-space graph search.
- Solves 16-bottle worst-case in under 2 seconds via DFS/BFS + pruning + memoization.
- Real-time step-by-step Swing GUI; MVC design pattern.
Open Source
Contributions
C++ · Sep 2025 · Upstream Contributor
OpenCV — Bug Investigation & Precision Fix
Diagnosed a floating-point precision anomaly in OpenCV's Windows build caused by MinGW FPU extended-precision behavior; implemented strict double-precision semantics via PR.