PORTFOLIO · 2026 · arsamsabbagh.ir
Arsam Sabbagh
آرسام صباغ
Senior AI Researcher · LLM Engineer · Web Designer (Coaxys)
Building, evaluating, and shipping intelligent systems at industrial scale — with a parallel practice in editorial web design under the name Coaxys. Based in Tehran, Iran.
me@asabbagh.com
About
Arsam Sabbagh is a senior AI researcher and LLM engineer with three-plus years building machine-learning systems for production. Alongside that work he publishes original research and long-form writing on machine learning for the Farsi-speaking tech community.
His core work sits at the intersection of LLM pipeline design, inference optimization, and quantitative evaluation of frontier models. Outside the lab he runs Coaxys, a small practice for frontend and brand work serving startups and tech companies.
- Senior AI Researcher · LLM Engineer
- Data-centric AI · agentic systems · inference
- Frontend & brand under Coaxys
- Go-to Farsi-language resource for ML engineers
Operating Principles
I. Evidence over enthusiasm.
Benchmarks are not vibes. Ship the eval before you ship the model.
II. Boring infrastructure, ambitious models.
A model is only as good as the pipeline that feeds it and the harness that scores it.
III. Persian is a first-class citizen.
A tokenizer that drops half the alphabet is a bug, not a quirk. Build for the language you write in.
IV. Editorial taste is technical taste.
How a project reads — its docs, its UI, its naming — is how it survives the year.
Expertise
Track A — AI & LLM Engineering
- Large language model pipeline design
- Inference optimization & throughput tuning
- Quantitative evaluation of frontier models
- Data engineering at industrial scale
- Agentic systems & deployment
- Applied ML research
Track B — Web Development & Design (as Coaxys)
- Frontend architecture (React, Next.js)
- UI/UX design & rapid prototyping
- Brand identity for tech products
- Editorial type & layout systems
Selected Projects
Frontier-Eval
A reproducible harness for quantitative evaluation of frontier LLMs across reasoning, code, and long-context benchmarks.
PythonvLLMEval
Pars-RAG
A Persian-first retrieval-augmented generation stack with domain adapters for legal and medical corpora.
RAGPersian NLPFAISS
Inference Lab
Throughput and latency tuning toolkit — quantization, speculative decoding, and KV-cache strategies benchmarked side by side.
vLLMTensorRTQuantization
AgentScript
A lightweight planner-executor pattern for production agents, with structured tools and replayable traces.
AgentsTypeScriptOpenTelemetry
Writing
28 long-form pieces on machine learning, LLM engineering, and applied research — read by the Farsi-speaking tech community.
- On Quantitative Evaluation of Frontier Models — 18,420 views
- Building a Persian-First RAG Stack — 12,110 views
- Speculative Decoding, Honestly — 9,870 views
- Agentic Systems Without the Cargo Cult — 14,230 views
- Inference Optimization: A Field Guide — 16,708 views
- Notes from the Frontier Lab — 8,990 views
Stats
- 28 articles published
- 118,280 total views
- 4,789 reactions
Current 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 (Coaxys): Next.js, React 19, Tailwind, Framer Motion, Figma, MDX
Contact
Open to senior ML engineering roles, research collaborations, editorial commissions, and selective design engagements through Coaxys.
آرسام صباغ — پژوهشگر ارشد هوش مصنوعی و مهندس مدلهای زبانی از تهران، ایران.
طراح وب تحت نام کواکسیس
تخصص: طراحی خط لولهٔ مدلهای زبانی، بهینهسازی استنتاج، ارزیابی کمّی مدلهای پیشرو، پردازش زبان فارسی