How I Built a Video Streaming Service
Turning raw uploads into adaptive, quality-adjusting playback with FFmpeg, without slowing the app down.
Software Engineer · Nepal
SAHILVERSE
Building the infrastructure that powers tomorrow.
Enter the universe↓Origin
I'm Sahil Dahal, a software engineer from Nepal. Self-taught since 2021, shipping to production since 2024.
I work across the full stack: the interface people see, the server behind it, and the systems keeping it running. When a problem demands depth, I go deep.
I don't build demos. I build software that runs: a open source package with 4,200+ installs, a platform people use daily, a deepfake detector that hits 95% accuracy.
The work spans web apps, payment infrastructure, developer platforms, and machine learning. Built across Node, Django, PyTorch, and .NET. Shipped to production.
Expansion
It wasn't planned.
It accreted.
Building since 2021 · Shipped to production
YearProject
2024
Full-stack cinema booking platform with interactive seat maps, eSewa payment integration, and QR-coded tickets. Prisma transactions ensure concurrent double-booking is structurally impossible.
2025
A programming language with Nepali-keyword syntax, built from scratch in TypeScript. Custom lexer, recursive-descent parser, tree-walking interpreter, and CLI with REPL, AST viewer, and token inspector. Published as a global npm package.
2025
eSewa Python SDKlive
Open-source Python package for integrating Nepal's primary payment gateway into any application. Framework-agnostic. Published to PyPI with 4,200+ total downloads and 350+ monthly installs.
2025–26
Deviolive
Full-scale developer learning platform. Isolated multi-language code execution, GPU-accelerated video transcoding via BullMQ, on-demand CTF lab orchestration, real-time collaboration, and gamification. Deployed on a self-managed VPS across 11 sprints. 78% of surveyed users said they'd use it.
2026
Multimodal deepfake detection: spatial CNN (EfficientNet-B2) and frequency-domain CNN (FFT + ResNet-18) fused through a custom Cross-Attention architecture with residual connections and Layer Norm. 95% accuracy, 0.99 AUC on FaceForensics++.
2026
Cross-platform journaling application built with .NET MAUI and Blazor Hybrid. Runs natively on Windows, macOS, iOS, and Android, with mood tracking, tagging, and analytics.
Client
Designed and built the complete website for Truth Advisor Education Foundation. Live and in production.
Client
Oscar Khadkalive
Designed and built a live personal website for Advocate Oscar Khadka.
Transmissions
Broadcasts, outward.
Deep dives into how I actually built these systems: the decisions, the tradeoffs, the numbers. First broadcast soon.
How I Built a Deepfake Detector
Two neural networks, one answer: spotting AI-faked faces in video, 95% of the time.
Inside Devio: Running Code Safely
How Devio runs a stranger's code in isolation, without putting the system at risk.
Composition
Three backend ecosystems.
Database to screen.
Down to the metal.
Languages
- TypeScript
- Python
- C#
- Java
- C++
Frontend
- React
- Next.js
- React Native
- Tailwind
Backend
- Node
- Express
- NestJS
- Django REST
- ASP.NET Core
Systems
- Docker
- Redis
- BullMQ
- MinIO
- FFmpeg
- Nginx
- Linux
Machine Learning
- PyTorch
- OpenCV
- EfficientNet
- ResNet
Signal