Armin StamateEmail me

working notes, november 2026

Los Angeles, California

Armin Stamate

the short version

I build private AI that runs on the hardware you already own. I study Mathematics of Computation at UCLA and founded NoemaAI, where I ship Noema, Noema Atlas, and the open-weight Noema 1.5 2B.

Streamed locally, 46 tokens at 36 tok/s, no network calls

keep scrolling

Education

currently studying

Mathematics of Computation

B.S., University of California, Los Angeles

When
Fall 2026 to Summer 2030
Where
Los Angeles, California
Now
Year one of four
20262027202820292030math I knowyou are here

Projects

Assistant for Apple platformssince 2025

Noema

A free, open-source AI assistant for Apple devices. It runs models, searches files, and uses tools locally.

I founded NoemaAI in July 2025, launched Noema 1 that September, and have continued developing the product through Noema 4.0 across local runtimes, private retrieval, model discovery, and tool use.

  1. One model libraryFive local backends: GGUF, MLX, ExecuTorch, Core ML, and Apple Foundation Models.
  2. Grounded in your filesConversations can use a personal dataset of PDF, EPUB, text, and structured files.
  3. Retrieval on the deviceSearch, OCR repair, and recovery paths run locally before the model answers.
  4. Live runtime statusWhich model is loaded and how much of its context the conversation is using.
Noema model library on Mac beside an iPhone answering a question with smart retrieval over a Linear Algebra dataset
fig. 1Noema 4.0 on Mac and iPhone. Hover a note to find it on the screen.

what I built

  1. On-device document retrieval

    Built private RAG for PDF, EPUB, text, and structured files, including OCR repair and retrieval recovery paths.

  2. Local API and tool execution

    Built an OpenAI-compatible local API, structured tool calling, offline Python, and voice input.

  3. Runtime and memory systems

    Integrated five local runtimes with device-aware memory admission, cache controls, and fit-aware model gating.

  4. Cross-device and release engineering

    Shipped cross-device inference, an 11-language native interface, and automated Apple build and release workflows.

Model distributionjuly 2026

Noema Atlas

A desktop app for downloading, verifying, and sharing local AI model weights across multiple sources.

Atlas combines resumable multi-source transfers, content-addressed storage, signed manifests, and direct handoff to LM Studio and Ollama.

fig. 2Every eligible route can contribute bytes to a transfer. Nothing enters the library until the complete file verifies.

what I built

  1. BitTorrent transport

    Built µTP and DHT transfers, discovery, concurrent downloading, seeding, and verified handoff into Atlas storage.

  2. Iroh multi-peer fetching

    Added striped downloads across multiple Iroh peers, live upload accounting, and fast peer-state invalidation.

  3. Transfer control

    Added live peers, upload ratios, recheck, sequential mode, queue priority, and scheduled bandwidth limits.

  4. Signed releases

    Built signed automatic updates and release paths for the native and Studio apps across desktop platforms.

Open-weight modelaugust 2026

Noema 1.5 2B

An open-weight 2B model that improves strict instruction following and multi-turn consistency while using fewer thinking tokens.

I developed Noema 1.5 from the earlier Noema 2B checkpoint using restoration-aware on-policy distillation. The release improves strict instruction following and complete multi-turn responses while using fewer thinking tokens than stock Qwen3.5-2B in the paired final evaluation.

IFEval, prompt strict

Stock 65.06, Noema 72.09.+7.02*

Multi-IF, complete three-turn

Stock 14.77, Noema 17.88.+3.11*

HumanEval+, pass@1

Stock 50.61, Noema 53.66.+3.05

Thinking tokens, relative to stock

−13.5%on average

fig. 3Paired final results against stock Qwen3.5-2B, with identical prompts, settings, and graders. Asterisks mark statistically significant gains. GPQA-Diamond and IFBench scored below stock.

Work experience

engineering and public service

  1. CSW69 delegationmarch 2025
  2. Eileen Higgins mayoral campaignjune to august 2025
  3. NoemaAIjuly 2025 to present
fig. 4The same work, laid out in time. Select a row to jump to the entry.

1.

july 2025 to present

NoemaAI

Founder and engineer

Founded NoemaAI; launched Noema 1 in September 2025 and led continuous development through the current Noema 4.0, integrating five on-device model runtimes, private document retrieval, tool calling, an OpenAI-compatible local API, voice input, and cross-device inference.

Los Angeles, California

Notes

Work on NoemaAI began in July 2025. Noema 1 launched in September 2025, and I have continued developing the product through the current Noema 4.0. I lead product engineering across native interfaces, model runtimes, retrieval, tools, testing, and Apple build and release workflows.

2.

june to august 2025

Eileen Higgins mayoral campaign

Field team volunteer, Office of Miami-Dade Commissioner Eileen Higgins

Served on the field team for ten weeks, canvassing Miami neighborhoods, phonebanking, identifying supporters, and answering residents' questions about the campaign platform.

Miami, Florida

Notes

I contributed to field operations during Eileen Higgins's successful campaign for Mayor of Miami. The work centered on direct conversations with residents and the practical discipline needed to turn a campaign plan into consistent field execution.

Responsibilities

  • Canvassed neighborhoods across the city.
  • Phonebanked to identify supporters and answer questions.
  • Communicated the campaign's positions on transit, affordability, and quality of life.
  • Followed field scripts, turf assignments, and daily campaign updates.

3.

march 2025

CSW69 delegation

Delegate, Mar Adentro de México A.C.

Participated with Mar Adentro de México A.C. during the sixty-ninth session of the UN Commission on the Status of Women at United Nations Headquarters.

New York, New York

Notes

CSW69 took place at United Nations Headquarters from March 10 to March 21, 2025. The session reviewed progress on the Beijing Declaration and Platform for Action thirty years after its adoption.

I participated with Mar Adentro de México A.C. and gained direct exposure to the pace, procedure, and follow-through of multilateral work.