Abozar Alizadeh · Software Engineer and Solution Architect

01 / Now

By night, it's one human and a crew of AI agents.

After work I build products on my own. My only team is a group of AI agents that I direct from a single control chat.

They research, build, review each other's work and write the docs. I decide what to build, test everything on my own phone and own every commit. It turns out one person can ship a lot this way.

This story runs backwards: the newest work first, then how I got here.

Put the AI only where the uncertainty is. Everything else should be boring code you can test.

Public GitHub contributions per year 2026 counted up to 14 October

203

757

github.com/abozaralizadeh

Public GitHub contributions per year
Year	Contributions
2020	55
2021	53
2022	10
2023	59
2024	74
2025	203
2026	757

How I build

01
Build, review, ship.
One agent builds a feature, another reviews it like a tough critic, a third polishes it. All before I install anything.
02
Measure everything.
Every feature ships with automated tests, and every claim is checked against real data before it goes out.
03
One session per project.
Every product keeps its own long-lived agent session with its whole history. One Mission Control chat steers them all.
04
Frugal by design.
No team to pay, no ad budget. Free cloud tiers, a Claude subscription and a lot of tests.
05
Build in public.
Public repos, a live trading dashboard and write-ups on Medium. Anyone can watch the work as it happens.
06
Ship the story too.
Every product ships with a build diary, demo films and social posts, made mostly by my own agents and skills.

contributions in the last 12 months: 809
public commits: 943
AI experiments running daily: 6
articles written: 20

02 / October 2026 · Coming soon

Something I can't show you yet.

03 / October 2026 · CronAI

Say it, don't cron it.

CronAI · open source

A neural network inside a TypeScript file.

A plain-English schedule picker for any website. A tiny 75k-parameter model lives in the source code and only fixes typos. A strict compiler writes the cron line, so it can't make anything up. On a typo benchmark, exact matches go from 16.9% to 93.8%. It also runs the schedules in AISM².

Code Widget builder The article

live · type a schedule, typos welcome

every weekday at 9am → 0 9 * * 1-5

04 / September 2026 · Two products, one night

One night in September, I started two products a minute apart.

Two agent sessions, opened sixty seconds apart: a booking platform for clinicians and a marketplace for group games. Three weeks later both were running, each with its own tests, legal research, films and build diary.

Praxamo · live at praxamo.it

Praxamo: your practice, online.

Praxamo banner: The patients you bring stay yours. A doctor by a sunny window, a phone with a booking on the desk — Praxamo posters and stills from its film · fictional practices, sample data, AI-generated photos

Four practice websites for a GP, a cardiologist, a dermatologist and a physiotherapist — Praxamo posters and stills from its film · fictional practices, sample data, AI-generated photos

Their own site, their own name.

Every Italian clinician gets a professional website and online booking at their-name.praxamo.it. No marketplace in between: the patients you bring stay yours.

A dentist, a GP, a cardiologist.

Eight templates and eleven colour schemes, all contrast tested, so each practice gets a site that actually looks like theirs.

Regulation is the architecture.

Every page is checked against Italian health advertising law before it can go live. Write "the best practice in Milan, guaranteed results!" and it says no.

Book without an account.

Patients book as guests on their phone. Confirmations and reminders arrive in the practice's own name and colour.

A human looks first.

Practices run their agenda, patients and availability from one console. And nobody goes live until a human (me) has checked them against the official registers.

Tested like it matters.

1,504 automated tests and 77 tenant isolation checks. Even the marketing film is a test: if the booking doesn't confirm while it records, the recording stops.

Visit praxamo.it

The film uses an AI voice and sample data.

Partyta · in demo, payments simulated

Partyta: any group game, one seat at a time.

Paintball, laser tag, softair and VR need 10 to 24 players. Nobody sells those games seat by seat, so you end up chasing friends for days.

On Partyta, venues publish open games and you join alone or with friends. Your card is only held, and it's charged only when the game reaches its minimum. If it never fills, nobody pays.

The logo is the product: a ring with one free place, and a coral dot filling the gap.

7of 24

Needs 3 more to confirm

Try it. The game is confirmed at 10 players.

Partyta player feed with VR and five-a-side games that need more players

A Partyta game page showing the card is only held until the game is confirmed

Partyta post: You don't need ten friends to play. A paintball group laughing in the woods, with a spare coral mask on a cable drum

Visit partyta.com

Promo with a synthetic voice. Photos are AI-generated.

demo-recorder · a Claude Code skill · open source, MIT

And a camera crew to film them.

Two days later I wanted demo videos of both products, so I built a skill for it. Ask Claude for a demo and it scripts it, records it with a smooth, visible cursor and captions, checks the result and hands you an MP4.

The Praxamo and Partyta films above were made with it.

› /plugin marketplace add abozaralizadeh/ClaudeScreenRecordingSkill
› /plugin install demo-recorder@abozaralizadeh

Code on GitHub

05 / July 2026 · AISM²

AISM² · open source

An agent that runs social media.

Write a brief and a schedule. When it's time, an agent reads the brief and its own memory, researches the topic, makes an image or a short Sora 2 video, writes the caption and posts to Instagram, X, YouTube or TikTok. Or it waits for approval.

It runs in production every day and manages my own X, Instagram and YouTube. One example is @genaicomicbook, the Instagram page of the ComicBook: AISM² posts every new episode and reel there by itself.

Schedules are written in plain English with my own CronAI widget, and for every instruction I choose how much freedom the agent gets: post on its own, wait for my approval, or just rehearse in dry-run mode.

automated tests: 2,242
live instructions: 21

Code on GitHub

AISM² banner: AI Social Media Manager, smarter content, stronger connections

The AISM² dashboard overview: 8 connected accounts, 15 of 21 instructions active, 459 posts published in 30 days at 97% success, performance, activity and upcoming posts

06 / December 2025 · trAIde

trAIde · open source · built in public

AI trading agents, with real money on the line.

A research agent scans about 700 futures contracts and the news. A trading agent decides. A supervisor talks to me on Telegram. And plain code, not the model, owns the stops.

Since 30 September I've been running an experiment: Jev vs Luna. Jev, a second decision model, trades next to Luna, the original agent, on the same account and behind the same safety gates. Each is scored only on its own calls, and the public dashboard compares their decisions side by side.

Code enforces survival. The model owns opportunity.

The public trAIde dashboard: win rate, indexed return, a one month equity curve, a win and loss ring and recent trade outcomes — The public dashboard at sandboxes.live/traide, one month view. Returns are indexed and normalized: it shows how the agents do, never the balance.

// built in public

Every decision, research note and supervisor note shows up live on a public dashboard.

// wide awake

A research agent scans about 700 futures contracts and keeps a calendar of big economic releases, stepping aside when they land.

// tested

1,360 automated tests guard the code that sits between the model and the money.

// principle

Decisions follow the market. Nothing is hand-tuned to this week.

Live dashboard Code on GitHub The full story

An experiment, not financial advice.

07 / 2025 · Small, sharp tools

Small, sharp tools I give away.

// pipup · PyPI + Homebrew

requirements.txt that tells the truth.

$ pip install requp
$ pipup
# pins requirements.txt to what's really
# installed. Never adds anything new.

Fun fact: the name pipup was taken on PyPI by a 2016 package, so it ships as requp. Both commands work.

Code PyPI

// Azure DevOps Marketplace

Gen AI Advanced PR Reviewer

AI code review on every pull request: inline comments, a security scan, and an approve or request-changes call. 60 installs, rated 5.0.

Marketplace Code

// PyPI + Homebrew

UPPPDF

Unlocks password-protected PDFs where the encryption allows it, trying several engines with fallbacks.

Code PyPI

×2 Pull Shark

// merged upstream

Contributions merged upstream.

microsoft/genaiops-promptflow-template a Publish Flow feature, merged in 2025
langchain-ai/langsmith-sdk a contribution, merged in 2024

08 / Since November 2024 · SandBox

AI experiments, shipped daily.

sandboxes.live is a playground of autonomous agents that publish something new every day, all by themselves: a comic that draws itself, tomorrow's newspaper, a blog that researches AI, and a mathematician that never gives up.

AI-generated comic art from the ComicBook

The daily lineup

trAIde dashboard

A public window into my trading agents.

Live decisions, research notes and supervisor notes from trAIde. Right now it also runs an experiment: Jev vs Luna. Jev, a second decision model, trades next to Luna on the same account and behind the same safety gates, and the dashboard compares their calls side by side.

Watch it live

ComicBook

A comic that draws itself.

One episode a day, in English, Italian and Persian, also posted to Instagram as @genaicomicbook. Keep scrolling: it has the best stories.

See the styles

AI Open Problem Solver

Chipping away at the hardest problems in math.

An agent works on the Riemann Hypothesis, P vs NP and the twin primes a little every day, with a memory of its dead ends and experiments.

Its own estimate of progress on Riemann

0.7%

Open the notebook

AI Blog

A blog that researches AI by itself.

Every day an agent searches the newest papers and lab announcements, reads them in a headless browser and publishes an illustrated post. It became my own morning reading.

Today's post

Tomorrow News front page for 15 October 2026 in English, marked as a fictional scenario

The same day's front page written in Persian

The same day's front page written in Italian

Tomorrow News

Tomorrow's newspaper, written today.

It reads today's real news and writes tomorrow's front page in English, Persian and Italian, each language written on its own. 580+ days of archive.

Read tomorrow

GenBox's retro TV in the desert, with an AI-generated news anchor in front of a world map

GenBox · where SandBox began

An AI government on a retro TV.

An imaginary AI world government researches the real world, makes one decision for humanity, then produces its own news bulletin with Sora 2. You turn the knob to change the day.

Tune in

Back in time

Two years of daily output, one calendar away.

Every experiment keeps its whole archive, and each page has a calendar to travel back. I went back to the first days and put them next to today.

GenBox's first broadcast: green text on a retro CRT in the desert — Captured from sandboxes.live on 14 October 2026 · AI-generated content

Tomorrow News in June 2025: a clean newspaper page with a lead story about ceasefire talks — Captured from sandboxes.live on 14 October 2026 · AI-generated content

GenBox · 8 Dec 2024 → Oct 2026

From green text to a news anchor.

The very first broadcast, on 8 December 2024, was green text scrolling on a CRT. Today the same imaginary government reads its decision on air, with an anchor generated by Sora 2.

Tomorrow News · 2 Jun 2025 → 13 Oct 2026

From a newspaper to a broadsheet.

By mid 2025 it already looked like a paper. Today it's a full broadsheet, written separately in English, Persian and Italian, with a red banner on every issue saying it's fiction.

AI Blog · 20 Apr 2025 → 12 Oct 2026

From big questions to arXiv deep dives.

The first post asked whether generative AI erodes our critical thinking. Eighteen months later the agent reads fresh preprints and explains them, every morning.

ComicBook · 5 May 2026 → 14 Oct 2026

24 arcs, 24 different artists.

The first arc, The Map in the Teacup, started on 5 May 2026 in a city that folds itself into new shapes every dawn. Twenty-three arcs later, every story still gets its own look.

ComicBook · sandboxes.live/comicbook · @genaicomicbook

Every story arc looks like a different artist drew it.

A chain of agents makes one episode a day, and every arc gets a new look. A Director plans the arc, a Storyteller writes, a Cartoonist draws, and native authors retell it in Italian and Persian without ever seeing the English. For every arc an Art Director researches a real art tradition, and a blind auditor checks the result without knowing what was intended.

Retold, not translated

Three languages, three native authors.

The Italian and Persian authors never see the English. They write from a shared beat sheet, so every edition reads like an original, and the Persian one runs right to left.

The Museum of False Winters, an arc written in English — English · The Museum of False Winters

Le lettere del ballo del disgelo, an arc in its Italian edition — Italiano · Le lettere del ballo del disgelo

The Department of Very Small Thunder in its Persian edition, laid out right to left — فارسی · اداره تندرهای خیلی کوچک

Three different arcs from 2026, captured from sandboxes.live/comicbook. Comic art is AI-generated.

Letters at Low Orbit: pixel art of a lunar mail station with Earth in the window — 5 Oct 2026AI-generated

The Hut Above the Vanishing Road: a woman watching a monsoon over salt flats — 5 Oct 2026AI-generated

5 Oct 2026 · Letters at Low Orbit

Pixel art on the Moon.

A quiet night shift at a lunar mail station, until a capsule arrives from a route everyone thinks is dead.

30 Sep 2026 · The Hut Above the Vanishing Road

A monsoon over the salt flats.

Rukmini watches floodwater threaten the only road home across Gujarat's Little Rann.

23 Sep 2026 · The Procession of Borrowed Faces

Masks in the drowned tunnels.

A city that survives in flooded subway tunnels, carrying masks that hold shared memories.

17 Sep 2026 · The Little Theater of Larkspur Quay

A storybook harbour town.

Mara inherits her mother's sign-painting shop, and the sealed little theatre hidden behind it.

11 Sep 2026 · The Third Beat Broadcast

Golden-age radio drama.

The national time signal has always had two beats. Tonight someone hears a faint third one.

4 Sep 2026 · The Bridge That Writes Names

Bazaar chromolithograph.

A festival around a living-root bridge, and a bridge that starts releasing names.

09 / 2020 to 2023 · Early open source

Before agents, there was plumbing.

My personal open-source projects from before the agent era: chatbots, phone calls and monitoring, built on my own time and shared with everyone.

// .NET · telephony

AudioSocket.Net

2023. Streams live phone-call audio out of Asterisk into .NET, the plumbing behind real-time voice bots.

Code

// Docker · 2,400+ pulls

offline-directline

2022. A Bot Framework Direct Line emulator I revised and shipped in Docker, so chatbots can be tested without the cloud.

Code

// Docker · 4,790 pulls

ApiWatchTower

2021. My first public tool. Calls your APIs on a schedule and alerts on downtime or when the response payload quietly changes.

Code

// npm · merged upstream

async-lock

2021. I added a maxOccupationTime option to this widely used Node.js lock library. It shipped in version 1.3.0.

Upstream

Docker pulls: 10.7k
public repos: 34
PyPI packages: 2
Marketplace rating: 5.0

Docker Hub All repos on GitHub

10 / Writing

I write about what I build.

20 articles on Medium, six of them published by Towards AI: generative AI in production, AI agents, and the experiments you just scrolled through.

All articles on Medium

11 / About me · 2020 to now

By day, I put generative AI into production.

At Vodafone Italia I grew from Senior Software Engineer (2020) to Technical Lead (2022) to Solution Architect and Team Leader (2023). Today I work as a Lead Architect on AI agents and LLMOps.

The project I'm proudest of: bringing generative AI into TOBi, Vodafone's virtual assistant, and into the tools the human agents use behind it. I led it from the first idea all the way to production.

// architecture

GenAI that scales

Azure Machine Learning, Prompt Flow and LangChain, designed for real traffic. Microsoft's advisors called it a new best practice for enterprise GenAI.

// upstream

Shaping the tools

I worked directly with Microsoft product teams. My feedback and code went into Azure ML networking, Prompt Flow and the GenAIOps templates.

// llmops

From notebook to prod

Pipelines that take prompts and models from an experiment to production safely, again and again.

// people

A team of nine

I lead and mentor an international team of nine engineers, local, remote and consultants, and I help hire the next ones.

years in software: 10+
engineers led: 9
clouds: Azure, AWS, GCP: 3
Kubernetes, certified: CKAD

Azure MLPrompt FlowLangChainLangGraphOpenAI Agents SDKAKSAPI ManagementFront DoorCosmos DBKubernetesTerraformPythonC# / .NETDjangoFlaskRedisPostgreSQLGCPAWS

12 / About me · rewind to 2006

Before Milan: three cities, three languages, one habit of making things.

2016 to 2019 · Turin

R&D, chatbots before they were cool.

Before Vodafone, I was in Turin. At BYTERS I built a Wit.ai chatbot for Yamaha Motor Italy, an Android NDK app that tags objects in live video with OpenCV, and an OCR receipt reader. Then, as Solution Architect at Djungle, I designed data pipelines and gamified products on Google Cloud.

2013 to 2016 · Turin

A new country, a new alphabet.

I had moved to Italy for my Master's at Politecnico di Torino. My thesis measured the distance between Android phones using Wi-Fi Direct.

Master of ScienceComputer Engineering and Communication NetworksPolitecnico di Torino · 2013 to 2016

2006 to 2012 · Hamedan

Learning how machines talk.

Before that, my Bachelor's at Bu-Ali Sina University, with a thesis that simulated backoff algorithms in ad hoc networks. On the side I supported live camera image processing software at Famori.

Bachelor of ScienceComputer Engineering (Software)Bu-Ali Sina University · 2006 to 2012

2006 · Tehran

And at the very beginning, I started in 3D.

My first job was modelling and animating for children's programmes at IRIB, Iran's national TV. For six years I also taught 3ds Max, V-Ray, Flash and C++ to anyone who wanted to learn. That's where this story starts.