AlphaGet in touch

Alpha Romer Coma

Machine learning engineer working with on-device LLMs and AI accelerators.

See the workGet in touch

Affiliated with

Projects.

KV cache prefix reuse, Experimental Machines

A phone that skips re-reading a 2,050-token agent prompt.

Built twice, because the app's two runtimes have different cache constraints.

27×

faster prefill on a second turn: Qwen3-1.7B on ExecuTorch, reusing 1,192 tokens

  • First turn5,852 ms
  • Second turn215 ms

  • OpenWeights, Experimental Machines

    Find a model. Run it here.

    Run open-weight models from Hugging Face on your Android phone, with llama.cpp or ExecuTorch.

    Two OpenWeights screens: Discover, a Hugging Face search with recommended ExecuTorch builds; and a chat with LFM2.5 1.2B that searches the web to answer whether to pack an umbrella in Tokyo tomorrow

    2.3×

    faster per token than llama.cpp on a Tensor G5, with LFM2.5 1.2B

    • Tensor G5, ExecuTorch46 ms
    • Tensor G5, llama.cpp106 ms
    • Snapdragon 8 Elite, ExecuTorch18 ms
    • Snapdragon 8 Elite, llama.cpp27 ms

  • ExecuServe, Experimental Machines

    Your phone is the server.

    An OpenAI- and Anthropic-compatible server for compiled ExecuTorch models, running in the background on Android.

    Two ExecuServe screens: Hosting, showing the server running with qwen3-1.7b as a hosted model; and Chat, with a reply from qwen3-1.7b and its prefill and decode speed

    0.41 s

    to first token on the NPU, with Qwen3 1.7B on a Snapdragon 8 Elite Gen 5 and a ~700-token prompt

    • NPU0.41 s
    • GPU1.0 s
    • CPU2.0 s

  • 18 models

    ExecuPack

    209 builds for phone CPUs, GPUs, and NPUs.

  • #23401

    PyTorch ExecuTorch

    Merged fix: LFM2 state no longer carries into a new conversation.

  • 4 setups

    De-mystifying PyTorch for ASICs

    When moving PyTorch off GPUs is worth it: 1× H100, 8× H100, TPU v6e-8, Trainium1.

  • 2,000 videos

    Micromarc

    The multimodal dataset behind Visual-Qwen, which flags cognitively-degrading short-form video.

Other projects

  • ExecuBench

    Benchmarks for ExecuTorch models on real phones, in progress.

  • Will It ASIC

    Can you run that model on that AI ASIC?

  • Compute Visualizer

    Why is this LLM workload slow?

Research.

What we measure at Experimental Machines, and on which hardware.

All research

9 of 9

models were fastest with NPU prefill

2 to 4 November 2026

Prefill and decode on CPU, GPU and NPU

Which processor should prefill and which should decode: nine models, every pairing, on a Snapdragon 8 Elite Gen 5.

  • 42%

    of 24 real tasks with Qwen3.6 on an RTX 4090, against 92% on Copilot Pro+ Auto

    • Pro+, Auto92%
    • Gemma 4, 409058%
    • Qwen3.6, 409042%

    GGUF, October 2026

    When is local Copilot worth it?

  • 63 to 71 ms

    per token on Inferentia2 from 1 to 32 streams, while throughput grows 26×

    • 1 stream63 ms
    • 8 streams65 ms
    • 32 streams71 ms

    ASICs, July and August 2026

    Llama 3.1 8B on AWS Trainium and Inferentia

  • 1.84×

    the H200's KV cache on an MI300X with 70 GB of weights loaded, from 1.36× the memory

    • MI300X102.25 GiB
    • H20055.60 GiB

    Hardware, July 2026

    MI300X against H200

Speaking.

From PyTorch Conference Europe to community stages across the Philippines.

All 23 talks

4,480

attendees across 27 sessions

Sessions by year

  • 22024
  • 142025
  • 112026

Experience.

Certifications.

Show 20 more

What people I've worked with say

What distinguishes Alpha further is his remarkable humility and positive attitude.

Despite his impressive credentials and achievements, he remains grounded and approachable, always willing to learn and grow. His blend of intellectual curiosity, leadership qualities, and character make him a standout individual.

Justine Jude PuraCS Faculty, FEU Tech

Data.

Scrapers and datasets.

  • TypeScript JobSpy

    Aggregates jobs from popular job boards with one call.

    A job scraping library for JavaScript and TypeScript.

    TypeScript code calling scrapeJobs for Indeed and LinkedIn, beside the jobs.json it returns

    ~1,067 jobs/min

    on Indeed, and ~157 on LinkedIn, measured from a residential IP on 31 August 2026

    • Indeed~1,067 jobs/min
    • LinkedIn~157 jobs/min

  • Philippine Mall Explorer

    What is inside Philippine malls.

    A reproducible dataset of what is inside Philippine malls, and a site for exploring it.

    Philippine Mall Explorer: a map of the Philippines with clusters of malls, and a list of the properties with the most listings

    42,313

    store listings across 297 properties and 10 operators, in the 2 November 2026 snapshot

    • Greenhills921 listings
    • SM City North Edsa868 listings
    • Festival Mall784 listings
    • SM City Fairview739 listings

Writing.

Notes on machine learning, software, university, and travel.

All posts

Get in touch.

For roles, research, and speaking.