Experience

Founder

ClassRec · classrec.com

Jan 2026 – Present
  • Built and deployed a full-stack web app enabling students to upload or live-record lectures and receive AI-generated transcripts
  • Eliminated ~30% of hallucinated transcript segments by integrating Silero VAD via ONNX Runtime for silence detection before passing audio to Whisper
  • Deployed faster-whisper large-v3 on Modal serverless GPU (T4), cutting per-lecture inference cost ~65% vs OpenAI Whisper API while improving accuracy on accented speech

AI Engineer

TalentSathi (Freelance) · Kathmandu, Nepal · Remote

Feb 2024 – Present
  • Built a classifier that flags bot and fraudulent applications before they reach human reviewers, filtering spam out of the recruiter queue
  • Automated first-pass candidate screening, cutting review time from 1 hour to 10 minutes per 100 applications

Technical Solutions Engineer

NestNepal

2026 – Present
  • Scoped and sized cloud infrastructure for customer workloads, translating traffic, storage and uptime requirements into concrete server, storage and network configurations

Education

B.S. Mathematics / Computer Science

St. Joseph's University, New York — Brooklyn · Minor in Data Science · GPA 3.6 / 4.0

May 2026

Projects

ColdVPN & ColdSpot

Your own one-click VPN

2026
  • Zero-config VPN: one download sets up a private server and starts encrypting all device traffic automatically
  • Captures traffic at the network layer, so even apps that bypass proxy settings are tunneled; ColdSpot relays it through a paired iPhone to a self-hosted exit server

CloudBot

AI food ordering chatbot

  • AI chatbot that takes food orders in natural language and processes them in real time
  • NLP intent recognition with Dialogflow ES; FastAPI webhook persists orders to Cloud SQL

MBGPT

Fine-tuned LLM for YouTube comment replies

2024
  • Fine-tuned LLM that generates context-aware replies to YouTube comments
  • Quantized Mistral-7B with QLoRA, cutting model size ~75% and enabling CPU inference on a 16GB MacBook
  • Parameter-efficient fine-tuning (QLoRA/PEFT) on a custom YouTube comment dataset