                    ⢠⣿⣿⣦⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣴⣿⣦⡀⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⠀⠀⢠⣿⣿⣿⣿⣆⠀⠀⠀⠀⠀⠀⠀⠀⣾⣿⣿⣿⣷⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⠀⢀⣾⣿⣿⣿⣿⣿⡆⠀⠀⠀⠀⠀⠀⣸⣿⣿⣿⣿⣿⡆⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⠀⣾⣿⣿⣿⣿⣿⣿⣿⡀⠀⠀⠀⠀⢀⣿⣿⣿⣿⣿⣿⣿⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⢸⣿⣿⣿⣿⣿⣿⣿⣿⣧⠀⠀⠀⠀⣼⣿⣿⣿⣿⣿⣿⣿⡇⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣠⣤⣤⣼⣿⣿⣿⣿⣿⣿⣿⣿⣷⠀⠀⠀⠀⠀
            ⠀⠀⠀⢀⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⠀⠀⠀⠀⠀
            ⠀⠀⠀⢸⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⠀⠀⠀⠀⠀
            ⠀⠀⠀⠘⣿⣿⣿⣿⠟⠁⠀⠀⠀⠹⣿⣿⣿⣿⣿⠟⠁⠀⠀⠹⣿⣿⡿⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⣿⣿⣿⡇⠀⠀⠀⢼⣿⠀⢿⣿⣿⣿⣿⠀⣾⣷⠀⠀⢿⣿⣷⠀⠀⠀⠀⠀
            ⠀⠀⠀⢠⣿⣿⣿⣷⡀⠀⠀⠈⠋⢀⣿⣿⣿⣿⣿⡀⠙⠋⠀⢀⣾⣿⣿⠀⠀⠀⠀⠀
            ⢀⣀⣀⣀⣿⣿⣿⣿⣿⣶⣶⣶⣶⣿⣿⣿⣿⣾⣿⣷⣦⣤⣴⣿⣿⣿⣿⣤⠤⢤⣤⡄
            ⠈⠉⠉⢉⣙⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣇⣀⣀⣀⡀⠀
            ⠐⠚⠋⠉⢀⣬⡿⢿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡿⣥⣀⡀⠈⠀⠈⠛
            ⠀⠀⠴⠚⠉⠀⠀⠀⠉⠛⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡿⠛⠋⠁⠀⠀⠀⠉⠛⠢⠀⠀
            ⠀⠀⠀⠀⠀⠀⠀⠀⠀⣸⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⠀⠀⠀⠀⣰⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣧⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⠀⠀⠀⢠⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⠀⠀⢠⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⠀⢠⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⢠⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣧⠀⠀⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⠀⣾⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡄⠀⠀⠀⠀⠀⠀⠀⠀
            ⠀⠀⠀⢸⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡇⠀⠀⠀⠀⠀⠀⠀⠀

NIR / Ly

I build high-performance backend systems, ML infrastructure, and production SaaS.

~/Contact

Email: [email protected] GitHub Matrix Discord: n.ir

Production

Remitae (B2B SaaS)

The Product: Automated invoice reminders and Xero synchronization platform. Built on Next.js with strict multi-tenancy via Postgres RLS.
Engineering Challenge: Architected a serverless job system (Vercel Cron) for 15-minute sweep cycles. Currently migrating from polling-based architecture to event-driven row triggers to eliminate race conditions in high-volume syncs.
Stack: Next.js • Supabase • OAuth2 • Stripe • Playwright • Zod
Status: Live Service

Systems & ML

Osu!Mania ML Benchmark

The Pipeline: End-to-end rhythm game automation pipeline. Built a 51GB ingestion pipeline and pivoted from API inference to a local Causal TCN. Refactored dataset storage from JSON to memory-mapped arrays to handle 26M+ frames with zero-copy loading.
Stack: Python 3.11+ • PyTorch • librosa • ossapi • Memory-Mapped I/O
Documentation: Read the Engineering Devlog

Neural Network From Scratch (C++)

Scope: Built a feedforward neural network entirely in standard C++ with a contiguous-memory tensor engine, dense layers (Xavier init), activation dispatch via function pointers, explicit forward-pass caching for backprop, and a modular SGD optimizer.
Key Work: Debugged silent indexing corruption from flat tensors, implemented mathematically correct gradient propagation validated through finite-difference gradient checking, and created an MNIST CSV pipeline with normalization, one-hot encoding, batching, and deterministic shuffling.
Links: Code • Devlog
Tech: C++ • Custom Tensor Engine • Gradient Checking • SGD • MNIST

TraceX (OSINT Engine)

The Engineering: High-throughput Rust ingestion engine for OSINT datasets. Engineered to be highly concurrent, utilizing Kafka for backpressure to handle spikes in data flow.
The Bottleneck: Ingestion speed consistently outpaced upstream API limits (Telegram), resulting in only 40% local CPU utilization. outcome: Project concluded after analysis showed that storage costs for the resulting dataset exceeded the project's economic viability.
Stack: Rust • Kafka • Elasticsearch • PostgreSQL • Docker

nirOS (x86-64 Kernel)

The Core: Custom x86-64 kernel implementing bitmap physical allocation, four-level paging, and a minimal VFS (tarfs, devfs).
Tech: C • x86-64 Assembly • Limine Boot Protocol
Source: github.com/nir659/nirOS

Less

Self-Hosted Infrastructure

Scope: Full mail stack (Postfix/Dovecot/DKIM), Multi-tenant hosting control planes, and Nextcloud instances. Focused on strict auditability, containerization (Docker/Borg), and cost-management.