Carlos Sanoja
Pixel-art night scene: a person at a desk watching a loss curve fall to 0.947 and a third-place mark appear, next to a Trainium rack with blinking lights, an ARC puzzle being solved, a robot arm and a drone on a shelf, a chalkboard, and El Avila mountain through the window.

I train language models against the clock.

I’m Carlos Sanoja, an ML research engineer. In 2026 I placed 3rd in Phase 1 of the AWS Trainium Frontier Competition, pretraining a model from scratch on one Trainium2 chip in 30 minutes, and I’m now in its 10-team final phase.

I also build agent systems that stop and hand control back to a person when the evidence runs out, and I write up what did not work as carefully as what did. I founded Quanta Labs, which builds AI tools for schools in Latin America, and I teach at Universidad Monteávila. Before language models I worked on robotics: SLAM, 3D perception and reinforcement learning for manipulators.

Read the write-upsAbout meGitHub

Selected work

Each write-up states the problem, the method and the results with their evaluation setting, and has a section on what did not work.

Aug to Oct 2026 3rd place, Phase 1

Pretraining against the clock on one Trainium2 chip

AWS Trainium Frontier Competition. Train a language model from scratch in 30 minutes of wall-clock on a single Trainium2 chip, scored on held-out bits per byte. I treated it as a measurement problem: profile first, change one thing, climb a validation ladder before paying for a full run, and pre-register the rule that decides.

Official val bpb 0.947182, down from 1.1728 over 16 submissions

Official Phase 1 scores Sixteen official submissions between September 3 and September 30, 2026. Validation bits per byte fell from 1.1728 to 0.947182. Lower is better. 0.951.001.051.101.15 Sep 3Sep 10Sep 17Sep 24Sep 30 0.947182, final 3rd 1.1728, 1st on day one
Every official submission in Phase 1, with the running best. Lower is better.
  • Oct 2026 Ongoing

    Trainium Frontier, Phase 2

    Sixteen Trainium2 chips, four hours per run, scored half on bits per byte and half on the CORE benchmark. The write-up comes after the deadline on November 4.

    Trainium2Neuron SDKNKIFP8
    5th of 10 finalists on the public board, Oct 19
  • Aug to Sep 2026

    COUNTERSIGN: an agent fleet that cannot sign

    Agents that do vendor onboarding and invoice verification end to end, where signing and paying are capabilities no agent holds. Every risk claim has to cite an evidence span or the verdict is rejected.

    Gemini on Vertex AIPythonAgent safety
    Signature denied in 12/12 runs; soak-test false positives cut from 80.7% to 28.1%
  • Aug 2026

    GradeSync: grading handwritten exams until a human is needed

    A fleet of Gemini agents on Google Cloud that grades handwritten exams from upload to gradebook, red-teams its own inputs, and holds a paper for the teacher when it should not decide alone.

    Google ADKCloud RunFirestore
    96 of 108 exams written to the gradebook unattended, 12 held for the teacher
  • Sep 2026

    Repaso: learning support with memory you can inspect

    A Telegram tutor on Strands Agents and Amazon Bedrock AgentCore that remembers which explanations worked for a learner, and a dashboard that shows the evidence behind every memory.

    Strands AgentsBedrock AgentCoreAWS CDK
    Two upstream fixes to the Strands SDK; AWS Builder Center article
  • Aug to Sep 2026

    Tracking dividing cells in zebrafish embryos

    Kaggle competition on 3D-plus-time microscopy. Oracle experiments showed linking was not the bottleneck, which redirected the work to detection and divisions.

    PyTorchILPKaggle
    Public leaderboard 0.608 to 0.946

Also competing

  • ARC Prize 2026, ARC-AGI-3Running and ablating a public agent harness for unseen interactive games on an open model. So far the clearest finding is that scores on the 25 public games overstate the leaderboard by about three times.
  • ARC Prize 2026, ARC-AGI-2Measured whether open teacher models could write solver programs to use as training data. The best one solved 2 of 35 evaluation tasks, so I paused the line.
  • RSNA Knee Abnormality DetectionTwelve-label knee MRI classification on Kaggle TPUs, with only 58 gold-labelled studies.
  • OpenCV AI Competition 2026Estimating how orbital debris tumbles from monocular video, with a system that declines to answer when the geometry does not allow it.

Open source

Earlier: robotics and perception

2017 to 2019, mostly ROS and Gazebo. Each project ships with a written report.

  • RTAB-Map SLAM and AMCL localizationMapping and particle-filter localization for mobile robots.
  • Follow-me droneA fully convolutional network that segments and follows a person in simulation.
  • ACLAHEA paper implementation of entropy-based adaptive contrast enhancement.