What I built, how I built it, and what it achieved.
Job seekers struggle to get realistic interview practice — AI can fill this gap with real-time, conversational mock interviews.
Built an AI-powered mock interview platform combining speech-to-speech AI (ElevenLabs) with expert-led sessions. Users speak their answers and receive immediate AI-driven evaluation and verbal feedback.
Real-time voice AI · LLM-based answer evaluation · GCP + Firebase backend
Traditional expense apps require navigating UIs — what if you could manage expenses through natural language with an AI assistant?
Built an expense tracking system using Model Context Protocol (MCP), enabling Claude to manage and interact with expense data via natural language. Eliminates the need for a traditional UI entirely.
Natural language expense management · Claude Desktop integration · FastMCP Cloud deployment
An AI agent that answers any question about you on your behalf, using real-time web data and context-aware intelligence. Acts as a personal representative powered by live information.
Multi-agent architecture · Real-time web data via Firecrawl · Text-to-speech responses
A multilingual voice translation system that converts spoken input into accurate translations across multiple languages using speech recognition and synthesis.
Multi-language support · OpenAI Whisper STT · End-to-end voice pipeline
Provides personalized food recommendations in under 5 seconds through a fast, voice-driven filter system without relying on traditional chatbots.
<5s recommendations · Voice AI · AWS Amplify deployment
Reproducing DeepSeek-R1-style reasoning capabilities in an open-source model requires careful RL fine-tuning — most tutorials skip the hard parts.
Fine-tuned Llama 3.1 8B for math reasoning using GRPO + LoRA, with a custom reward function based on symbolic equivalence checking. Enabled consumer-GPU training via 4-bit quantization.
78.5% GSM8K accuracy · LoRA + GRPO pipeline · Consumer GPU via 4-bit QLoRA
Choosing the right inference framework can 2-3x your throughput — but most comparisons are shallow and don't test real-world conditions.
Benchmarked LLM inference performance by running three models across vLLM and SGLang on AWS EC2 with GPU-accelerated Docker. Measured throughput, latency, and resource utilization under controlled load.
3 models benchmarked · AWS EC2 GPU deployment · Throughput & latency analysis
Students spend hours creating study materials manually — AI can generate personalized quizzes, mind maps, and flashcards instantly from their notes.
Platform using multi-model RAG to help students prepare for exams with personalized study materials, practice questions, mind maps, and more.
Multi-Model RAG · Qdrant vector DB · Hackathon winner on Devpost
Docker documentation is dense and abstract — most people learn best by seeing and interacting, not just reading.
A web-based learning game with 15 interactive, animated levels teaching Docker from "What is Docker?" to real-world deployment. Built with vibe-coding using Claude Code.
15 interactive levels · Animated visualizations · Terminal-style demos
Writing a tailored cover letter for each job posting takes 10+ minutes — a browser extension can do it instantly from the job description and your resume.
Browser extension that instantly generates tailored cover letters from any job posting using your resume. Saves ~10 minutes per application, built with vibe-coding.
One-click generation · Resume-aware · ~10 min saved per application