Ahmed Sozzer

austin, tx

open to full-time from

Ahmed Sozzer — Full Stack Engineer

I write the layer most people import.

Full stack engineer at FYCLabs. I sit with the client, find the step a person is doing by hand, and build whatever it takes to delete it — the product in Next.js, React Native and Postgres on GCP, and the systems underneath it in C++ and Python when the product needs them.

started PlusWeb because I wanted to know what Express was actually doing.

plusweb — 0.0.0.0:3000 C++17 
$ cmake --build build -j && ./plusweb
listening — one libuv loop, keep-alive on

GET    /users/new        200   0.041 ms
GET    /users/123        200   0.038 ms   :id = "123"
POST   /users            201   0.039 ms
GET    /admin/metrics    404   0.036 ms

three dependencies, coffee

See the work

projects

Three things I built. Two run in your browser, one runs a printer.

  • a web framework
  • a 3D printer protocol
  • a convnet in a canvas

01

PlusWeb

An Express-style HTTP framework in C++17.

Everything above the socket is mine: the accept loop, the parser, the router, the middleware machinery. A literal path segment always beats a parameter, whatever order you registered them in.

  • C++17
  • libuv
  • llhttp
  • CMake
  • MIT
dependencies
three — libuv, llhttp, nlohmann/json
tested
GoogleTest, ASan + UBSan in CI
license
MIT, installable via CMake
curl — localhost:3000 keep-alive 
$ curl -i http://localhost:3000/users/123
HTTP/1.1 200 OK
Connection: keep-alive

{ "id": "123", "name": "Ahmed" }

$ curl -i http://localhost:3000/admin/x
HTTP/1.1 404 Not Found

the timing on that page is measured in your browser, on a port of the same trie — not typed in by me.

Open the playground repo

02

AMS-X

An open modular filament system for Bambu Lab printers.

Stock hardware caps multi-material at four spools and asks a human to stand there for every change. None of the protocol is published, so the command set was pulled apart against a live machine.

  • Python
  • FastAPI
  • MQTT
  • ESP32
  • 3MF
plan source
the sliced .3mf, parsed server-side
transport
MQTT over TLS, FTPS on LAN
spool module
human now, TMC2209 next
device/+/report mqtt · tls 8883 
{ "print": {
    "gcode_state": "PAUSE",
    "layer_num":   48,
    "sequence_id": "2041"
} }

retract → select → feed → sense → resume

nothing moves until the printer’s own report agrees with the plan.

How it works repo

03

MNIST

Convolutional digit recognition, served as a static file.

Two convolution and pooling stages learn the strokes, then a dense head turns them into ten probabilities. Trained with a library, served without one — the weights ship as a plain JSON file and the forward pass is written out by hand in the page. No TensorFlow.js, no ONNX runtime, no WASM blob. Nothing you draw leaves your machine.

  • Python
  • TensorFlow
  • CNN
dataset
MNIST, 60,000 / 10,000
shape
conv 32 → pool → conv 64 → pool → dense
serving
static hosting, inference client-side
in the browser
no runtime — the forward pass is hand-written

and no, numpy doesn’t count — there isn’t any in a browser anyway.

Draw a digit repo

Experience

Three roles since 2024, all of them client-facing. I scope with the people paying for it, build it, and own the pipeline that ships it — frontend, backend and CI/CD, usually all three on the same project.

every role, in detail

  1. FYCLabs

    Full Stack Engineer

    to now

    I scope with client founders and end users directly, then ship it — web and mobile products on Postgres and GCP, and the release pipelines that carry them.

    Two things recur: integrations against GoHighLevel, Zoho and HubSpot that delete an operations team’s manual data entry, and CI/CD with test gates running across several concurrent client projects.

    • Next.js
    • React
    • TypeScript
    • PostgreSQL + Prisma
    • GCP
    • React Native

    7× the users on the same backend — query and schema work, then scaled out

  2. Holiday Channel

    Full Stack & iOS

    to

    An e-commerce platform end to end, built so the checkout path holds on the busiest week of the year.

    Checkout through order fulfillment, with the service architecture designed to absorb holiday traffic without degrading the one path that is not allowed to degrade. Google Ads on the acquisition side.

    • Next.js
    • Shadcn
    • Framer Motion
    • iOS

    shipped alongside my last semester at Illinois

  3. Luminii LLC

    Data Analyst & Software Intern

    to

    An ML pricing system on live supplier data, replacing a re-calibration somebody was doing by hand.

    The pipeline sat on a database that had grown past its indexes, so that came with it — a reindex and automated validation underneath, which is what actually made the query times usable.

    • Python
    • SQL
    • ML

    a manual re-calibration process, gone

Education

A CS degree, and the research that came out of it.

University of Illinois Urbana-Champaign

B.S. Computer Science

Before any of the jobs: a matching algorithm for pairing mentors with mentees at scale. The work is credited in a work-in-progress paper (PDF) for the American Society for Engineering Education, and I presented it twice.

The tech stack

most days it is the top row; the rest of the list is what the problem asked for

languages
  • TypeScript
  • JavaScript
  • C / C++17
  • Python
  • Java
the C++ one
systems
  • POSIX sockets
  • MQTT
  • multithreading
  • CMake
  • ESP32
  • Linux
AMS-X
web
  • React
  • Next.js
  • TypeScript
  • Node
  • Express
  • FastAPI
the client work
mobile
  • React Native
  • Expo
  • EAS builds
  • Firebase Auth
the client work
testing & profiling
  • GoogleTest
  • ASan / UBSan
  • perf
  • load generation
the write-up
data & cloud
  • PostgreSQL
  • Prisma
  • Hasura GraphQL
  • MongoDB
  • Redis
  • GCP
  • Firebase
  • Docker
the client work
models
  • TensorFlow
  • NumPy
  • a forward pass written out by hand
MNIST