Pats Laurel

I am a machine learning engineer
at Thinking Machines.

An NVIDIA B200 GPU taken apart: the package substrate, the interposer, two compute dies with the L2 cache between their halves, eight HBM3E memory stacks, and one streaming multiprocessor lifted out and opened to show its four tensor cores.
Plate 1 One decode step on an NVIDIA B200 (HGX); the SM is enlarged. Measured: 148 SMs (Chips and Cheese, U. Delaware); about 21 TB/s within a half (Chips and Cheese). Reported: 8 × 24 GB stacks (Tom's Hardware, SemiAnalysis). 4 tensor cores per SM (inferred from Chips and Cheese). Other numbers: NVIDIA.

Writing

Plates 2 to 4

A unit square and two images of it under a matrix: one stretched by the area scale factor |det(A)|, one mirrored.

Plate 3 · 10 min read

The Determinant: How Much of the Plane Survived?

I computed determinants for longer than I'd like to admit without knowing what the number meant. It measures area: how much a square grows or shrinks when a matrix moves the plane, and whether it comes back flipped. Once I saw that, I stopped needing to memorize the rules.

A grid turned about its origin, with the point (1, 1) carried to (0, 1.42).

Plate 4 · 8 min read

What Does a Matrix Actually Do?

A 2×2 matrix is four numbers, and those four numbers decide where every point in the plane goes. Drag the basis vectors and watch the grid rotate, stretch, shear, and flatten.

All writing, 12 pieces

Work

Chronological table, Sep 2022 to nowCurrentPast

  1. Machine Learning Engineer · Engineering Consultant

    Thinking Machines Data Science

    Sep 2025 to present · Philippines · Remote

    Client projects from the first brief through delivery: agent architecture on AWS, data-platform design, proposals and cost models, and the build itself. I also write our internal engineering standards, multi-cloud runbooks, and AI-assisted project templates that cut new-project setup from about two weeks to a day.

  2. Team Lead, Full Stack Engineer (AI)

    NuWorks Interactive Labs

    Mar 2024 to Sep 2025 · Full-time

    Led a small team shipping AI products end to end: model integrations, application code, deployments, and the production work after launch.

  3. Full Stack Engineer

    NuWorks Interactive Labs

    Oct 2023 to Feb 2024 · Part-time · Remote

    Joined part-time, shipped product features and model integrations, then moved into the lead role.

  4. Technical Writer

    Tutorials Dojo

    Oct 2023 to Feb 2024 · Contract

    Wrote AWS and machine-learning guides on SageMaker pipelines, serverless inference, and model fairness.

  5. Head of Data Science

    Short Stay AI

    Jun 2023 to Feb 2024 · London, UK · Remote

    Owned the models, AWS setup, and data pipelines for a small London startup.

  6. Data Science Mentor

    Eskwelabs

    Sep 2022 to Feb 2024 · Part-time

    Reviewed capstone code and helped project teams make better technical decisions.

Talks

I spoke at AWS Community Day Manila 2025 about building trustworthy AI.

All talks and slides

Conference

30 mins

AWS Community Day Manila 2025 · Metro Manila, Philippines

Is Your LLM Lying? A Practical Guide to Building Trustworthy AI

My AWS Community Day Manila 2025 talk on testing an LLM app before and after it ships: Bedrock evaluations, contextual grounding guardrails, prompt management, and logging.

LLMs · Evals · Guardrails · AWS Bedrock

Read the talk notesSlides (opens in a new tab)

Interests

Outside work, I'm heavily interested in inference and kernel programming.

For no good reason

History related to wars and East Asia history like the Chinese dynasties and Japan history like the shogunate, samurais, etc.

A katana taken apart along its tang. From the hilt end: tsuka, tsuba, habaki and the blade with its nakago, with the saya below the blade.
Plate 5 A katana taken apart along its tang, beside its scabbard.

Books

  • Man's Search for MeaningViktor Frankl
  • Machine Learning with Amazon SageMaker CookbookJoshua Arvin Lat
  • Machine Learning Engineering on AWSJoshua Arvin Lat

Contact

Stamp: the NVIDIA B200 board from Plate 1, drawn exploded

dev@patslaurel.com

Metro Manila, Philippines

GitHub (opens in a new tab)LinkedIn (opens in a new tab)

0/2000