Software engineer — systems & AI.
I build the layer that has to be fast and the AI that has to ship.

I'm an AI Software Engineer at DOMO — Uzbekistan's utility super-app (~2M users) — where I build the assistant and LLM-serving infrastructure behind the product. That work lives in a private GitLab, so what's public here is the systems side of how I think.

I write Python and Rust, and I went down to Mojo to build the project below.

A Telegram Bot framework written in pure Mojo. No Python in the runtime.

murtazox04/mojogram · docs · MIT

Mojo gives you static types, value semantics, and real parallelism with no GIL — but no Python stdlib to stand on. So the foundations are written from scratch:

• JSON — hand-written parser + serializer over an arena DOM. ~2.4 µs/op, ~420K updates/sec. Adversarial suite: emoji, surrogate pairs, scientific notation, deep nesting, unicode keys.
• HTTP — no native TLS in Mojo, so the wire is system curl via subprocess, with retry/backoff and 429 retry_after honored. ~2.3 ms/request floor — inside Telegram's ~33 ms budget.
• Concurrency — opt-in parallel batches via parallelize; shared FSM state behind an Atomic spinlock. Stress-tested at 500 concurrent writes + 1000 contended read-modify-writes, zero torn state. ~4× over sequential.
• Webhooks — pure-Mojo HTTP server over libc sockets via FFI, secret-token validated, tested end-to-end through a real ngrok TLS tunnel and scaled behind nginx.

~60 typed API methods, typed updates over the JSON DOM, composable filters, inline/reply keyboard builders, Payments + Telegram Stars, i18n with locale fallback, token-bucket rate limiting. CI runs every suite on each push — pixi + uv, macOS + Linux — tracking a pre-1.0 language.

Not a port. Where Mojo wouldn't express something — no heterogeneous function values, so no decorator handler registry — I redesigned around the language instead of fighting it.

• Production LLM systems — RAG, agentic loops, MCP servers, eval pipelines, cost-optimized inference for a high-traffic consumer app.
• Uzbek speech & language — synthetic Uzbek TTS and an Uzbek-LLM corpus pipeline. Low-resource-language work: most of the fight is data, not modeling.
• Systems performance — the thread behind mojogram and my Rust work. Close to the metal, where it pays.

Languages   Python · Rust · Mojo
Backend     FastAPI · Django/DRF · async · Taskiq / Celery
Data        PostgreSQL · pgvector · Redis
AI / ML     PyTorch · Hugging Face · vLLM · LoRA/QLoRA · RAG · MCP
Infra       Docker · Kubernetes · GitHub Actions · GCP

Telegram · LinkedIn · murtazox04@gmail.com