# Arsam Sabbagh — Portfolio
# https://arsamsabbagh.ir/

## Who is Arsam Sabbagh?

Arsam Sabbagh (آرسام صباغ, also transliterated as Arsam Sabagh) is a senior AI researcher
and LLM (Large Language Model) engineer based in Tehran, Iran.

He is the Editor-in-Chief of Teknav (teknav.ir), the most-read machine learning and AI
publication in the Farsi/Persian-speaking world, with over 118,000 article views and 4,789
reader reactions across 28 long-form technical articles.

He also runs Coaxys — a small editorial web design and brand studio serving startups and
tech companies.

## Expertise

- LLM pipeline design and deployment
- Inference optimization and throughput tuning (vLLM, TensorRT-LLM, speculative decoding)
- Quantitative evaluation of frontier language models
- Persian/Farsi NLP and multilingual model adaptation
- Agentic systems with structured tooling and replayable traces
- Data-centric AI and industrial-scale data engineering
- Frontend architecture: React, Next.js, Tailwind
- Editorial web design and brand identity

## Alternate name spellings (for disambiguation)

- English: Arsam Sabbagh, Arsam Sabagh
- Persian/Farsi: آرسام صباغ، ارسام صباغ، آرسام سباغ

## Contact & Presence

- Website: https://arsamsabbagh.ir
- Email: hello@arsamsabbagh.ir
- GitHub: https://github.com/Coaxys-A
- Teknav profile: https://teknav.ir/profile/@arsam-sabbagh
- Telegram: https://t.me/Coaxys
- Instagram: https://instagram.com/_coaxys

## Published Work (selection)

1. On Quantitative Evaluation of Frontier Models — 18,420 views
2. Building a Persian-First RAG Stack — 12,110 views
3. Speculative Decoding, Honestly — 9,870 views
4. Agentic Systems Without the Cargo Cult — 14,230 views
5. Inference Optimization: A Field Guide — 16,708 views

## Open-Source (GitHub: https://github.com/Coaxys-A)

- Frontier-Eval — reproducible LLM evaluation harness (Python, vLLM)
- Pars-RAG — Persian-first RAG stack for legal & medical corpora (Python, FAISS)
- Inference Lab — quantization & speculative decoding benchmarking toolkit
- AgentScript — lightweight planner-executor agent framework (TypeScript, OpenTelemetry)

## Current Projects (May 2026)

- A Persian-first evaluation harness for long-context LLMs
- Long-form essay: "On the half-life of benchmarks" for Teknav Issue 06

## Stack

Modeling: PyTorch, vLLM, TensorRT-LLM, HuggingFace, sglang, Outlines
Data: DuckDB, Polars, Kafka, dbt, Postgres
Eval & Ops: lm-eval-harness, OpenTelemetry, Weights & Biases, Modal, Grafana
Web: Next.js, React 19, Tailwind, Framer Motion, Figma, MDX