Skip to content
moe-babazadehckpt 2026.10

model · moe-babazadeh · trained 2007 → 2026

Moe Babazadeh

CTO at Neuroject. I architect AI-driven platforms and scalable digital products.

CTO at Neuroject, building the alignment layer for construction enterprises. Building since 2007: robots, IoT and firmware, then web, mobile and AI platforms.

Fig. 1 forward(x)

11 of 11 inputs on, 11 of 11 outputs fire.

x · input · click to ablate

ŷ · click to trace

Fig. 1. A forward pass from skills to what they shipped: 11 inputs, two layers of labelled features, and a sigmoid head with a bias that puts every output at p ≈ 0.95 when all inputs are on. The weights are set by hand, not learned. With no hidden biases the network is piecewise linear, so gradient × input is exact here: the paths into an output sum to its logit minus its bias.

model.summary()
Layer (type)               Output shape   Param #   Non-zero
────────────────────────────────────────────────────────────
input (InputLayer)         (None, 11)       0         0
features_1 (Dense, relu)   (None, 9)      99        19
features_2 (Dense, relu)   (None, 7)      63        15
shipped (Dense, sigmoid)   (None, 11)     88        39
────────────────────────────────────────────────────────────
Total params: 250   Non-zero: 73 (71% sparse)
Head bias: [-5.54, -4.57, -4.57, -7.86, -6.65, -6.76, -3.40, -3.00, -3.00, -6.90, -9.08]

model card · v2026.10

Model card

Model details

I started building websites as a freelancer in 2007, studied electrical engineering with a focus on control systems, and spent my university years building robots, including a 1st place at RoboCup IranOpen 2010.

Since then I’ve built for an electronics and IoT company, a grocery marketplace, an edtech company and a construction-tech company. Today I’m CTO at Neuroject, leading the platform that connects BIM, schedules, budgets and risks for construction enterprises, and bringing AI into it.

Along the way I’ve shipped apps to the App Store and Google Play, hardware that’s still logging data after 15 years, and fixes to open source projects like Gatsby and DefinitelyTyped.

Intended use

  • Architecture, roadmap and delivery of cloud-native, multi-tenant platforms.
  • Bringing AI into products: natural-language access to project data, early warning on schedule and cost risk.
  • Hardware to cloud: firmware, boards, mobile apps and the platforms behind them.
  • Setting the engineering bar: a permission check on every API route, audit trails, alarms kept in infrastructure as code.

Training data

Pre-training on freelance web work from 2007, then electrical engineering, robots and a paper.

SplitDatasetNotes
2008–2012BSc Electrical Engineering, Control SystemsUniversity of Tabriz. Thesis: a two-wheel self-balancing robot with real-time fuzzy control.
2009–2011Small Size League soccer robotsA team of five fully autonomous robots playing a real match.
2010RoboCup IranOpen, 1st placeDeminer Technical Challenge.
2010National Skills Competitions, bronzeMobile Robotics.
2014–2016MSc Nanoelectronics EngineeringUniversity of Tabriz. Thesis: plasmonic wave sources based on multilayer structures.
2017Paper on arXiv ↗Design, engineering and optimization of a grid-tie multicell inverter for energy storage.

Input features

  • TypeScript
  • React & Next.js
  • Laravel (PHP)
  • Python & FastAPI
  • AWS
  • React Native & Expo
  • Embedded C
  • PostgreSQL & MySQL

Caveats and recommendations

Fig. 1’s weights are set by hand to make the attributions readable, and the loss in Fig. 2 is illustrative. The dates, products, devices, links and figures on this card are real. For a fuller eval, read the CV or ask the model directly.

scalars · train/loss

Training run, 2007 → 2026

Runs

Each run is fine-tuned from the checkpoint before it. Loss is illustrative; run dates are real.

train/loss

0.00.51.01.52.02.520072009201120132015201720192021202320252027nowFreelanceNiknam Electronic RayanseniorAlget.nl0&1NeurojectCTONeuroship
Fig. 2. One run per role, x in years. Circles mark where a run starts from its parent’s checkpoint; diamonds mark promotions inside a run.
  1. 2026-05_neuroshipChief Full Stack Engineer (contract) · NeuroshipMay 2026 – Present

    Contract role alongside Neuroject.

    • Chief Full Stack Engineer on contract, alongside my role at Neuroject.
  2. 2024-06_neurojectChief Full Stack Engineer → CTO · NeurojectJune 2024 – Present

    Built NeuroBuild, then CTO of the alignment layer for construction enterprises.

    • CTO since September 2026, after two years as Chief Full Stack Engineer. I lead the architecture, roadmap, and delivery of a cloud-native, multi-tenant platform that connects BIM, schedules, budgets, and risks into traceable decisions.
    • Bringing AI into the platform, from natural-language access to project data to early warning on schedule and cost risk.
    • Led the architecture and build of NeuroBuild on Next.js, React 19, and Laravel, hosted on AWS Amplify and Laravel Vapor, with 4,700+ commits across five codebases.
    • Set the engineering bar: a permission check on every API route, audit trails, and production alarms and GuardDuty alerts kept in infrastructure as code.
    neuroject.com ↗
  3. 2021-02_zerooneLead Frontend Developer · 0&1February 2021 – June 2024

    Active-Class LMS and proctoring, Attendum firmware and apps on both stores.

    • Led the frontend architecture of Active-Class, an integrated LMS and exam monitoring platform, including identity checks during live exams with the Microsoft Azure Face API.
    • Shipped Attendum end to end: firmware for the ESP32 card readers and nRF52840 BLE beacons, the iOS and Android app, the staff dashboard, and a Canvas LTI app.
    • Built Office.js add-ins for PowerPoint and Word, the School.how dashboard, the Education Guide app, and the Learning Kit and Exam Kit websites.
    • Redesigned the CI/CD pipelines to cut build and deployment overhead.
    0and1.co.uk ↗
  4. 2020-08_algetSenior Frontend Developer · Alget.nlAugust 2020 – February 2021

    A React marketplace that took local grocers online.

    • Built the React frontend of a digital marketplace that helps local supermarkets and grocery businesses sell online.
    • Implemented state management and client-side data handling that cut unnecessary API requests.
    • Shipped MVP features quickly with the product and backend teams, and improved deployment workflows with CI/CD.
  5. 2012-10_niknamSoftware → Senior Software Engineer · Niknam Electronic RayanOctober 2012 – August 2020

    IoT platforms, live GPS tracking and connected devices.

    • At Niknam Electronic Rayan, built IoT platforms and connected systems for monitoring, tracking, and data visualisation across multiple industries.
    • Designed web applications and real-time dashboards, and integrated hardware devices with GPS and mapping services for live tracking.
    • Modernised legacy platforms for performance, usability, and maintainability. Software Engineer from 2012, Senior Software Engineer from 2017.
  6. 2007-08_freelanceWeb Developer · FreelanceAugust 2007 – October 2012

    Websites and web apps for clients in PHP, JavaScript and jQuery.

    • Built websites and web applications for clients with PHP, JavaScript, jQuery, HTML, and CSS.

projector · pca

Everything shipped, embedded

  • Products at work
  • Hardware and research
  • Open source
  • Personal

31 points · 64 dims → 3 · perplexity 10, initialised from PCA · drag to rotate

2024 · Neuroject

NeuroBuild

The alignment layer for construction enterprises: one multi-tenant platform for BIM models, WBS, quantity take-off, bills of quantities, tenders, and 4D progress reports, now with AI for natural-language questions and early risk warnings. I lead it as CTO.

  • Next.js
  • React
  • Laravel
  • AWS

Nearest neighbours cosine

  1. 0.766
  2. 0.706
  3. 0.653
  4. 0.642
  5. 0.601

edge · 5 devices in production

Running at the edge

Not models: boards and firmware, in production. From GSM weather stations that have been logging for over 15 years to battery-powered card readers at universities in the UK and US.

  • LG-01 in production

    Logger

    GSM weather station

    Several sensors per station, with logs sent home over GSM.

    • 50+ stations
    • 15+ years online
    • 0.5B+ records
  • EPC-01 in production

    EPC

    Pressure and flow controller

    Embedded hardware and firmware that hold pressure and flow at their setpoints.

    • In production
  • PLC-01 in production

    Mini PLC

    Compact PLC

    Built on STM microcontrollers, for automation on a small footprint.

    • In production
  • AT-01 in production

    Attendum reader

    ESP32 card reader

    NFC, RFID, BLE and over-the-air updates. Wireless, with a battery that lasts about 16 months.

    • 200+ devices online
    • 3M+ records in 2 years
    • 10K+ classes covered
  • AT-02 in production

    Attendum beacon

    nRF52840 BLE beacon

    Bluetooth proximity check-in, updated over the air with DFU.

    • In production

evals · upstream

Open-source evals

Fixes contributed upstream to libraries that are downloaded more than 300k times a week, sorted by reach.

BenchmarkTaskPatchStarsWeekly downloads log
gatsbyjs/gatsby ↗split 2022SecurityRemoved the deprecated uuid@3.4.0 from gatsby-cli and gatsby-plugin-sharp.That version can fall back to Math.random(), which is unsafe for IDs. npm flagged it on every install.★ 56k163k
nanostores/persistent ↗split 2022Data lossSynced storage events only for the store’s own key.Opening the app in a second tab wiped saved data.★ 38382k
HuolalaTech/react-query-kit ↗split 2026Type safetyOpt-in strictVariables: typed query variables become required at the call site.A forgotten argument compiled fine and only failed at runtime.★ 48631k
DefinitelyTyped/DefinitelyTyped ↗split 2025Type safetyTyped the optional RegExp filters on the Forge Viewer’s getPropertyHashes.Calls that Autodesk documents failed to type-check.★ 51k22k
microapps/gatsby-plugin-react-i18next ↗split 2021BugAdded forwardRef support to the localized Link.Refs passed to localized links were silently dropped.★ 1195.4k

inference · mailto

Prompt me

user16 tokens · 55 chars

assistanttoy tokenizer · ids are hashed

Send opens your mail app with this message, addressed to moe@babazadeh.dev.

Send ⌘ ⏎

tools.json

{
  "email": "moe@babazadeh.dev",
  "linkedin": "/in/moe-babazadeh/",
  "github": "github.com/moeBabazadeh",
  "cv": "/resume.pdf"
}

Or call one of the tools directly.