My new professional page is live, and it explains all my projects in one place. By day I work on GenAI in production. By night I am one human with a crew of AI agents, building on my own. On the page you will find: Praxamo: websites and online booking for doctors, 1,504 automated tests Partyta: join a paintball or laser tag game one seat at a time (demo, payments simulated) AISM2: an AI agent that runs social media trAIde: trading agents with real money and a public dashboard, 1,360 automated tests Sandboxes: AI experiments shipped daily, 580+ days of Tomorrow News My public GitHub contributions: 203 in 2025, 757 in 2026 so far. Which project do you want a deep dive on first? abozaralizadeh.github.io

trAIde security fix: failed Telegram requests could put bot tokens in exception logs. I added redaction at every log handler, including tracebacks, with tests for token patterns and env secrets. github.com/abozaralizadeh/tr…

Read “17 Days with Jev: I Put a Sub-Second “System One” AI on a Real Trading Account“ by me on Medium: pub.towardsai.net/17-days-wi…

a crew of AI agents draws a new comic episode every day and nobody touches it. this is the cast of one of the story arcs the trick that keeps the art on style is a blind auditor. one agent describes the image without knowing what style it was supposed to be, a second one compares that description to the style card without ever seeing the image. so neither of them can just agree with what it was told to expect the chain is Director → Storyteller → Cartoonist → native retellers in Italian and Persian. handoffs are schema checked in plain code, the agents only do the creative part whole thing is open source, link below 👇

Anthropic's own numbers: junior tech hires dropped from 27% to 9% of their tech roles since early 2025. Engineers now commit 5x more code with agents. In my projects the agents do the typing and I review. More code from agents means more review for me, not less. And juniors always learned from the small tasks. Now agents do those tasks. So if nobody hires juniors today, who will review the agent code in 5 years?

Passed 250 followers! 🎉 Two weeks ago I had about 100. Today I have 257. I'm Abozar, Lead Architect in Milan. By day I build AI agents and LLMOps at work. At night I build my own projects alone, with AI agents. In this short time I found great people here. You gave me real ideas, like a waitlist before registration, a TTL for my agent lock, and logging every approval next to the proposal. Small things, but very useful. And thanks to you, my work now reaches many more people than I could reach alone. All my projects are on one page: abozaralizadeh.github.io Thank you to everyone who replied, asked hard questions, or just said hi. Next stop is 500. If you build things too, let's connect! 👋

Been running this for almost 2 years now: 6 AI agents that publish something new every day, and nobody touches them a comic that draws itself (new art style every arc), tomorrow's newspaper written today, an AI government that makes a decision and then reports on itself on its own TV news, a blog that reads new AI papers, a math open problem solver and two AI traders all open source. if you like it, a ⭐ on GitHub would honestly make my day github.com/abozaralizadeh/Sa… sandboxes.live

Partyta lets you join paintball or laser tag without a full group. Join an open slot seat by seat. Your card is held until quorum, so you pay only if the game runs. Payments are simulated. The real test is whether people come back. Would you try it? partyta.com

Hi Open Source Software! I'm Abozar, Lead Architect in Milan. At night I build with AI agents. Since November 2024 I run sandboxes.live. It's a small playground of autonomous agents. Every day they publish something new, all by themselves. What runs every day: - ComicBook: a comic that draws itself. A Director plans the story, a Storyteller writes, a Cartoonist draws, and a Reteller writes it again in Italian and Persian. Not translated, retold. - Tomorrow News: reads today's real news and writes tomorrow's front page in 3 languages. 580+ days of archive now. - AI Open Problem Solver: works a little every day on Riemann, P vs NP and twin primes. It has a Python math sandbox and remembers its dead ends. Its own estimate on Riemann is 0.7% done. Very honest agent. - AI Blog: finds new AI papers, reads them in a headless browser and writes a post. It became my morning reading. - GenBox: an AI world government makes one decision a day, then makes its own TV news with Sora 2. Two things I learned: - Parallel runs took the same job and burned tokens. A simple lock fixed it. - Round robin for LLM calls is not like HTTP. A Sora job must stay on the same resource until the end. All open source, MIT: github.com/abozaralizadeh/Sa… If you like it, a star on the repo helps me a lot. And tell me which agent you would read every day.

Claude Code Review is now GA for Team and Enterprise plans. Anthropic's numbers from June to September: merged PRs more than doubled, and the share of reverted changes fell by more than 40%. The second number is the interesting one. With agents, more code is easy. Less code to undo is the hard part. Quick, Standard or Deep review per repo. Which one would you pick for prod?

Hi Build in Public, Abozar from Milan here. Lead Architect by day, at night I build SaaS alone with AI agents. Working on now: Praxamo, websites and online booking for doctors. It has 1,504 tests and 0 users, because registration is not open yet. At least the tests are happy. Question for you: did you open signups before the product felt ready, or after?

Building Praxamo taught me consent is not a yes/no flag. When I added GA4 beside Clarity, an old “yes” could enable both. Now consent records which tools it covered, and asks again when that set changes. How do you handle this? praxamo.it

Hi AI community! I'm Abozar, a Lead Architect in Milan. My day job is GenAI and LLMOps, so I put AI into production. After work I build my own projects alone, with AI agents as my team. One of them is sandboxes.live, a hobby where autonomous agents run in daily loops. One agent there, Tomorrow News, has run for 580+ days. Running agents every day teaches you the boring lessons fast: a retry policy (mine stops after 3 tries), a load balancer in front of the models, and a lock so parallel loops do not burn tokens on the same job. All my projects are on one page: abozaralizadeh.github.io What is the hardest part of taking AI from demo to production for you?

The model behind your API name can change while you sleep. DeepSeek just shipped V4.1-Flash-1017. They say it is faster, finishes tasks quicker and gives shorter answers than the last version, so better quality and lower cost at the same time. Before the launch, the new id showed up in the API for a short time, and testers noticed the general deepseek-flash endpoint already giving shorter, clearer answers. Good news, mostly. But shorter outputs are also a change in behavior. If your agent parses the answer, expects a certain format, or your evals were tuned on the old version, a quiet update can break things that worked yesterday. The safer way: pin the dated model id in production, not the general alias. Put the new version behind the same load balancer, run it on real tasks, compare cost per finished task, then switch. Do you pin model versions in production, or ride the latest alias?