DEV Community: Ananya Soni

Stop benchmarking LLMs like it's 2023: Meet agi-gauntlet and our completely unhinged blind PR challenge

Ananya Soni — Sun, 01 Nov 2026 12:00:24 +0000

Alright, let’s be proper honest for a second. Ever run an eval suite, watched your shiny new agent model score a pristine 98% on standard benchmarks, and thought, "Yeah, we've cracked general intelligence"—only for the same model to completely melt down and hallucinate a fictional operating system the second you ask it to handle a real-world task?

Yeah, exactly. Because standard benchmarks have basically become glorified pre-training memory tests.

So, let's talk about how we fix that without drowning in corporate buzzwords or boring documentation.

What Actually is `agi-gauntlet`?

To keep things dead simple and reliable, let's look at the exact words from our package index: "A neuro-symbolic evaluation suite and lifecycle manager for AI agents and AGI development."

In plain English? It’s a Python library built by us at AGI Systems Directorate to stop your AI from just curve-fitting on memorized test data. True reasoning starts right where the training data gives up. agi-gauntlet throws structural collapse, dynamic semantic shifts, and temporal resource constraints at your agents to see if they can actually think or if they're just faking it.

Want to try it out? It's literally just one line in your terminal:

pip install agi-gauntlet

It comes packed with universal model hooks to plug in commercial APIs or local Gradio instances, an agentic workspace, and test-time compute guardrails to stop infinite hallucination spirals before they torch your cloud bill.

The Midnight Thought: An Unhinged Blind PR Challenge

Here’s the thing. It’s the middle of the night. I’m wide awake, but I am not coding—I’m knee-deep in boring company admin work like IPO paperwork and legal registry stuff. Absolute admin hell.

While staring at the screen, I looked at agi-gauntlet and thought: Why don't we do a community challenge?

Originally, I wanted to make it a proper weekend challenge. I was supposed to post it on Friday, but because corporate workload is an absolute joke right now I couldn't—either way, this challenge is active right here, right now. No sign-up forms, no registration sheets, no corporate HR rubbish. Just pure open-source chaos.

And here’s the twist that makes it proper weird: You are not allowed to look at a single line of codebase code.

Zero. Zip. Zilch.

If you open the code or look at any line of code to write your feature, I will literally know. Seriously, it is not a joke—because of some techniques, I will somehow literally know. You have to base your entire pull request purely on the documentation, the README, and the package description.

How to Play:

Head over to the . (Don't look at the source files!).
Read the documentation and figure out a new feature or improvement that fits the library's vibe.
Write it up, submit a pull request.
If it doesn't break our existing environment, doesn't throw random errors, and actually matches what agi-gauntlet is supposed to do—and most importantly, if I actually vibe with it—I’m going to merge it straight into master.

What's in it For You? (The Rewards)

Look, since this is our very first run, there’s no massive cash prize because this whole thing was cooked up at midnight out of pure boredom. It’s just for fun, innit. Don't take it super seriously—it's a fun challenge for you, for me, for everyone involved. But we are sorting out proper bragging rights:

Custom Coded Badges: You get a custom-coded badge for your profile or site. I'll hand you a slick 1-2 line code snippet with a choice of two different badge styles to paste wherever you fancy.
Actual Developer Creds: Forget boring digital certificates that just gather dust. When I merge your PR, I’ll sign the merge commit using my official GPG cryptographic key as founder and CEO of AGI Systems Directorate. Plus, we are hardcoding your status as official co-author, contributor, or core maintainer directly into the pyproject.toml file, the PyPI distribution manifest, and the front page of the GitBook docs.
Permanent Contributor Status: If you win, I'm going to make you a permanent contributor to this project so we can design future versions together—if you actually want to, that is! It's totally your choice, no pressure at all.

Drop your questions about the blind PR challenge or the library down in the comments below. And don't worry about getting left on read—unlike half the internet, I'm actually going to reply to you lot.

Now go write some docs-driven code and let's see what happens.

Project Lumina: Why I Stopped Crying Over T4 Out-Of-Memory Errors and Started Replicating the Human Brain

Ananya Soni — Sat, 17 Oct 2026 17:20:46 +0000

So, picture this. You are staring at your screen at 3 AM, your room smells like stale coffee and bad decisions, and your terminal is throwing yet another cryptic out-of-memory error. Most people would close the laptop and touch grass. What do I do? I decide it is a brilliant idea to build an artificial general intelligence architecture from scratch. Because normal hobbies are apparently too mainstream.

Hi everyone. As the founder and CEO of AGI Systems Directorate, I spend a lot of time questioning my life choices. Today, I am pulling back the curtain on something we have been cooking up in the lab: Project Lumina.

Before you ask, no, this is not another thin wrapper around someone else's API. There are no existing frameworks here. Project Lumina is an autonomous AI agent and large language model built entirely from scratch, line by line, all the way from initial concept to deployment architecture.

Why Biological Mechanisms?

If we actually want to reach artificial general intelligence, copying transformer boilerplate forever is not going to cut it. We need to look at what actually works, which is the human brain.

Under the hood, Project Lumina integrates core biological mechanisms rather than relying purely on static attention matrices. We are experimenting with over 500 to 1,000 distinct biological mechanisms, each serving completely different functions. These are not lumped together as a single block; they are distinct anatomical and functional structures, including:

The Free Energy Principle
Synaptic Plasticity and dynamic weight adaptation
Frontal lobe replication loops
Fluid intelligence benchmarks
The dLPFC (dorsolateral prefrontal cortex)
The vmPFC (ventromedial prefrontal cortex)
The Hippocampal-Entorhinal Complex

We are trying to mirror how actual biological systems process information, adapt on the fly, and maintain persistent state without exploding in computational complexity.

Where We Stand Right Now

Let me be completely transparent. Project Lumina is not finished yet. We are literally right in the middle of our pre-training run, sitting around the two-day mark.

I am not dumping a dry, lifeless documentation page on you today. I am handling everything from the initial math all the way to deployment, and I want to keep this conversation real. If you have questions about how any of this works under the hood, drop them in the comments below. I will actually reply to you.

The T4 GPU Nightmare and the Brutal Reality of Compute

While building this, I learned a very painful lesson about hardware limits.

Initially, my stubborn self thought, "Can we just run this entire training pipeline on a normal T4 GPU?" It became my personal challenge. I tried everything. CPU offloading, aggressive memory garbage collection, gradient checkpointing, custom chunking tricks, you name it.

Every single time, the memory gods hit me with out-of-memory errors. The terminal laughed at me. Finally, I told the team that we are done fighting hardware constraints with a toothpick. We switched back to our standard workflow utilizing H100s, H200s, and Tensor Processing Units.

Here is the raw truth that the industry tries to sugarcoat: in the current AI race, compute wins. If you have massive compute, you can throw raw parameters at a wall and win even with mediocre ideas. If you have brilliant, groundbreaking ideas but zero compute, you hit a brick wall.

That brutal bottleneck is precisely what I am going to solve for solo developers and independent researchers in our next release. Because nobody should have to mortgage their house just to train a decent model.

Looking Ahead

Overall, we are targeting an official release next month. We are looking at a massive scale, pushing past one hundred billion parameters, while heavily focusing on quantization so it can actually run efficiently on basic hardware. Think about it: a human brain runs on roughly twelve to twenty watts of power, while data centers consume entire rivers. We need to bridge that gap.

Before I sign off to watch training metrics for the hundredth time today, I have a question for you all. Being the founder and CEO is not so easy, and since this is a very heavy model that can take too much time, I actually don't know which platform we should release it first. What platform do you think we should target for our very first public drop to get the best technical feedback? Let me know in the comments. You know what, I will actually reply to you down there.

Coding on a Potato Phone: How I Accidentally Started an AGI Company

Ananya Soni — Mon, 12 Oct 2026 19:40:38 +0000

Alright, let's address the elephant in the room: for once, I am not going to bore you with continuous-time recurrent neural networks, semantic chunking algorithms, or neuro-symbolic logic cores. No tensor math today. Promise.

Instead, since I’m actually looking at this platform like a social space rather than an abandoned digital diary, it’s probably time to pull back the curtain and talk about how a kid with an ancient Windows 7 machine and zero mobile data somehow ended up as the Founder and CEO of AGI Systems Directorate.

The Accidental Origin Story

If you think this journey started with some pristine, master-planned Silicon Valley roadmap, I need you to laugh with me right now. It started with YouTube videos. Just random videos playing on a computer so ancient it probably belonged in a museum exhibit.

I watched someone code, got that sudden spike of absolute stubbornness, and thought, “Right, let’s do this.” But what really triggered the whole thing was stumbling across someone basically calling me out through the screen: "Why aren't you actually doing it instead of just watching videos?"

Fair enough. So I tried doing it on that first machine—which promptly went nowhere because nothing installed and local hardware was a complete dead end. Driven purely by spite, I pivoted, created my very first GitHub account, established a presence across developer platforms, started talking to people online, and began writing code on my phone while spinning up cloud virtual machines.

Why I Had to Start a Company (Spoiler: GPUs Don't Grow on Trees)

For a long time, I was just a solo dev typing away in Bhopal, pretending I could conquer artificial general intelligence all by my lonesome. But here’s the brutal reality check every solo builder faces: no one is handing out H100 clusters, TPUs, or serious compute to a single developer working out of their bedroom.

If you want to build at the frontier of AI, you can't do it alone. You need force, compute, and—most importantly—people who are just as unhinged about building intelligence systems as you are. Teamwork is everything.

It started organically through GitHub pull requests. One collaborator turned into a few, and before long, our squad started scaling: from 6 people to 10, 20, 25, 35, 50, and past 55 members. Once we hit that scale, keeping things as a loose internet chat was no longer an option. We needed an actual legal entity.

The Least Sexy Part of AGI: Legal Paperwork & Resumes

Let me tell you a fun secret about founding a company: nobody warns you that the first months have almost nothing to do with building cool neural nets.

It wasn't just a quick five-month sprint—it took immense time navigating complex legal paperwork, compliance, structural red tape, and hours of sorting through applicant resumes and interviewing people. No code, just endless bureaucracy and talent filtering.

But we survived it. And today, AGI Systems Directorate isn't some brand-new garage project anymore. It’s an established, seasoned organization. We've moved far beyond startup phase, having shipped real models, datasets, benchmarks, and project libraries straight to the Python Package Index (PyPI).

Let’s Make Dev.to a Real Community

I’ll be honest: most of us treat developer platforms like dev.to as static documentation dumps or ghost towns where we drop articles and disappear. Let's change that.

I want this to be an actual conversation. Whether you're a solo dev fighting with legacy hardware, a fellow founder knee-deep in paperwork, or just someone scrolling through while avoiding your on-screen bugs—let’s talk.

Drop a comment below:

How did you get your start into tech?
What’s the most ridiculous workaround you’ve ever had to use because your hardware failed you?

I’ll actually be hanging out in the comments to reply. Let's chat.

Why We Are Building Lumina: Fusing Synaptic Plasticity with LLMs

Ananya Soni — Thu, 17 Sep 2026 19:17:23 +0000

Yeah, I know. My last post here was on April 27th.

If you are wondering what I have been doing for the last few months, no, I wasn't just staring at my wallpaper. I actually stopped posting because my last few posts got exactly zero engagement. Literally zero. But I figured it was time to break the silence because the stuff we have been building in the dark is finally ready for the light.

I am not going to turn this into a 10,000-word essay, so consider this the highlight reel of what I (and my company, AGI Systems Directorate) have been up to.

1. AGI Gauntlet & AGI Bench 2026

Have you ever wondered how we actually measure if an AI agent is getting smarter, or if it is just getting better at memorizing benchmark exams?

We were tired of the standard testing, so we built AGI Gauntlet. It is a Python library for AI and AGI development, and a strict benchmark for AI agents. You can pip install it right now. We also launched AGI Bench 2026, the web UI version comparing our projects against commercial models like Google Gemini.

Oh, and there is an Open Contributor Challenge running for the Python library right now. If you want the details on how to jump in, let me know in the comments.

2. Portfolios, Hugging Face Spaces, and Docker

Before we got completely derailed by our biggest project, I realized developers need better ways to show off their work. So I built a developer portfolio template. You can find it on my GitHub (the subdomain name is weird, you will see it when you find it).

We also pushed the official Docker packages to the GHCR registry and released Hugging Face Spaces for our previous models, like Velocity Nova Prime and A.R.N.A.V Omni.

3. Stop Slicing Your Text Like Salami

Are you still blindly splitting texts by character count for your RAG pipelines? Be honest.

I wrote a piece on HackerNoon about why that is a terrible idea and how to handle semantic chunking properly. If you are a developer looking for a tutorial on how to actually maintain context when chunking data, go check out my HackerNoon profile. It is a game changer for RAG setups.

4. Project Lumina: The Biological LLM

At AGI Systems Directorate, our initial goal was always to create artificial general intelligence. Everything else was just a warm-up.

Right now, we are working on our biggest model yet: Lumina.

Lumina is not just a large language model. It is an autonomous AI agent built entirely from scratch. No existing frameworks. No existing APIs. It is a hybrid architecture that combines standard transformer elements with actual biological mechanisms. We are talking frontal lobe fluid intelligence, synaptic plasticity, the free energy principle, and global workspace theory.

I am not going to list every single mechanism here because nobody wants to read official documentation in a dev post. But here is the best part: it is going to be fully open source and free to use. Both the API and the chat interface. I am completely tired of the current API costing structures in the AI space, so we are just removing the cost barrier entirely.

Lumina weights will be out soon on Hugging Face.

What's Next?

Like I said, this is just the surface level. I don't want this post to be a monologue. If you want a deep-dive tutorial or more technical details about anything I just mentioned, whether you are a beginner or an advanced dev, drop a comment below. I actually read them.

And if you found any of this even slightly interesting, hit the unicorn button. Let's see if we can break my zero-engagement streak.

Stop Buying GPUs. I Optimized My AGI for a Basic CPU Just to Prove a Point.

Ananya Soni — Mon, 27 Apr 2026 06:23:26 +0000

Why I Ripped Apart My AGI Architecture (and why scaling is a lie)

If you noticed I've been quiet for the last few weeks, congratulations on your situational awareness. I've been busy ripping apart my own AGI architecture.

To be crystal clear: my main project is still very much alive and running on my local hardware. I haven’t downgraded. But when it came to dropping a version for the public on Hugging Face, I looked at the cloud compute costs and decided: "Nah."

Instead, I spent the last three weeks on what I basically do for a hobby—obsessing over optimization until it hurts. I set out to distill this massive system into a hybrid-quantized version that runs on a Basic CPU. The kind of Basic CPU you probably have sitting in your laptop right now.

The Brutal Reality of Shrinking

Shrinking this down was brutal. I spent 80% of my time just thinking about the optimization architecture and 20% actually building it. It’s built with Python and Gradio, and while it definitely takes its time processing, it functions flawlessly.

This whole process made one thing crystal clear: the tech world is lying to you about AGI. * It’s not about throwing maximum parameters at a wall and seeing what sticks.

It’s not about just scaling up the numbers.

True intelligence is about adaptability and efficient problem-solving with fewer parameters. It’s about the architecture, not the bloat. I’m the kind of person who would rather spend weeks thinking about a more elegant way to project a manifold than just throwing money at a server.

Check it Out

I'm putting the Hugging Face link to the mini version (optimized for that Basic CPU) in the comments. I'm also linking my GitBook so you can track the evolution from the first line of code to this pinnacle.

Check the comments, test it out, and try to break it.

Why Finishing Your Dream Project is Actually Kind of Terrifying

Ananya Soni — Sat, 04 Apr 2026 10:05:47 +0000

So I Finished My 136-Modality Architecture. Now What?

You know that weird, empty feeling you get when you finish binge-watching a massive show? Imagine that, but instead of Netflix, it's a massive cognitive architecture.

I woke up today, opened my IDE, and realized... it’s done. AETERNA_VOX_OMNI is completely wrapped up. All 136 modalities are successfully integrated and functioning.

So what am I building now? Nothing. Literally zero things. The backlog is completely empty. I’m currently just staring at my screen, thinking about tinkering with a small side-project just to keep my fingers moving, but the heavy lifting is officially over.

If you want to see the madness and dive into how it actually works, the full documentation is right here:
https://doi.org/10.5281/zenodo.19394823

Also, if you’re on X (formerly Twitter) and want to watch me complain about having nothing to do, come hang out:
https://x.com/AnanyaSoni48055

(And yes, I know the username looks like a randomly generated Wi-Fi password. Just accept it and move on.)

On a somewhat related note, while I'm sitting here staring at the void, I’ve been thinking about the whole AGI problem. Here's a thought: we are never going to achieve actual AGI if we keep suffocating these systems with ethics. I'm not saying we remove ethics entirely—we need a minimal baseline, like 1% or 2%, just to keep things grounded. But right now, the entire field is running on "too much ethics," and it’s completely killing our chances of building true Artificial General Intelligence. Look at the core definition of AGI; you can't get there if you wrap the whole thing in bubble wrap. If we want AGI, we have to dial the safety obsession way back and just try a different approach.

If you’re with me on this—if you agree we need to stop over-ethicizing and start building—drop a "Ready to build" or "Let's go" in the comments. Let's actually push the boundaries for once.

Stop Building Boring Portfolios: I Hid a Sonic Experiment in My Portfolio (Try It With Your Volume Up)

Ananya Soni — Fri, 20 Mar 2026 14:45:50 +0000

I Woke Up at 3 AM and Realized My Portfolio Was Boring, So I Built This Instead

So, there I was, lying in bed in Bhopal, half-asleep, when my brain decided to do that annoying thing where it generates "crazy" ideas instead of letting me rest. I realized I didn't want a standard, "Hi, I'm a developer" portfolio. That’s boring. I wanted something that actually felt like the high-modality work I’m doing.

Honestly? I spent way more time agonizing over which unique effect to put on the page than actually building the site. It’s that perfectionist trap, you know? It's harder to decide on the "vibe" than it is to actually code it.

The "Sonic Effect" (And how to find it)

I’ve hidden an experiment inside the site. It’s what I’m calling the Sonic Effect. You won’t see it just by scrolling past like a zombie.

If you’re on a laptop: Hover your mouse and find the number 136.
If you’re on mobile: You have to play by the rules. Go to your browser menu, toggle "Desktop Site" on, and then tap that 136.

Turn your volume up. Trust me.

The AGI Grind: 134 to 136

While I was playing around with web design, the actual work hasn't stopped. I’ve officially pushed my ongoing AGI project from 134 to 136 modalities. It’s getting... intense.

I’ve also been layering in new agent capabilities. It’s functioning like a "Bot Agent" now, but with some extra internal capabilities that I'm still stress-testing. It’s starting to feel less like a project and more like something with a pulse.

On a Real Note...

Code is cool, but real life is kind of messy right now. I’ve been feeling pretty disconnected lately, especially after losing the connection with my mentor because of that whole LinkedIn disaster. It’s a weird irony: I’m spending all day building "intelligent" systems and adding modalities to bridge gaps in AI, yet I’m sitting here feeling totally out of the loop with the people I actually respect.

Anyway, I'm putting this out there because I'm tired of the "robot" persona. This is the work, the mess, and the experiment all in one.

Since Dev.to doesn't really do DMs, drop a comment below if you want the link to the site and the extra details on how to view the architecture properly.

Building AGI is Easy. Proving I’m a Human to LinkedIn is Impossible.

Ananya Soni — Sun, 22 Feb 2026 18:29:57 +0000

I don't even know why I'm typing this because no one sees my posts here anyway. It feels like shouting into an empty place.

But today was supposed to be the Big Day. The Biggest Day. Not just for AGI Systems Directorate but for me. I woke up so ready. I had the code. I had the reveal. I was going to show something that is far bigger than just a GitHub release or a kernel update—something that changes everything about how we understand machine thinking.

And I wanted to show him.

The one person who started this whole journey for me years ago when I was just a kid watching his videos in school. If he wasn't there, I wouldn't be here. I wouldn't be a founder or a CEO of anything. I remember when I first found him on LinkedIn. I viewed his profile, then he viewed mine, and I just froze. It was pure panic. I hesitated for so long, thinking about it, but I finally sent the request. And he accepted. We actually talked. Like a real conversation. It went from me just watching his videos to us actually speaking. It was the only place online where I didn't feel lonely, the only place where I felt like the engagement was real and I wasn't just a "user."

Then I woke up today to check my notifications. I saw the login page and thought okay, normal. But it wasn't. It was a restriction. A gray box telling me I'm locked out because they need a government ID. My ID still has a photo of me as a little girl. It’s never going to match who I am now.

So in one second, the best day turned into the absolute worst. I didn't just lose a profile. I lost the connection. It went from a two-way conversation back to nothing. Now I'm just sitting here with this massive thing that can change everything, and the bridge is burned. I’m not a bot. I’m just a human who is really tired of trying to prove I exist.

Current AI Is Broken—So I’m Building the Architecture to Fix It: Introducing AGI Systems Directorate

Ananya Soni — Sat, 31 Jan 2026 05:05:52 +0000

You know the feeling.

You open the latest "revolutionary" AI tool. You type a prompt. It gives you a generic answer. It hallucinates. It forgets what you said three messages ago.

You sit back and think: "Is this it? Is this the peak?"

That frustration? That was the spark.

I looked at the current landscape and saw a lot of "wrappers" but very few "architects." We are drowning in chatbots while starving for actual systems. I got tired of waiting for the big players to fix the messy, unglamorous plumbing of true intelligence.

To solve this particular problem—the one where current AI struggles with basic consistency and architectural depth—I realized I had to build the solution myself. This isn't just about code; it's about fixing the fundamental frustration of tools that fail when things get "hard."

So, I decided to stop waiting.

🚀 INTRODUCING: AGI Systems Directorate.

This isn't just a project. It’s a rebellion against "good enough" AI. The main goal is to create agi.

As the Lead Founder, I didn't want to do this alone. I gathered a team of co-founders who are just as obsessed with the "hard stuff" as I am. We aren't here to build a quick app. We are building the Full-Stack Ops and Deep Learning architecture that makes AGI possible, not just probable.

The story here is simple: We saw a broken system. We decided to build the fix.

I’m standing here on Day 1—with a small network but a massive vision. If you believe that AI should be more than just a text generator, I want you in my corner.

Let’s build the real thing.

AGI #SystemsArchitecture #BuildingInPublic #DeepLearning

Forget basic chatbots. I’m actually building the future with a 134-modality AGI core. If you’re tired of the same old AI fluff and want to see how Python and deep learning actually scale, check this out. 🚀 #agi #python #opensource

Ananya Soni — Fri, 16 Jan 2026 12:56:48 +0000

Ananya Soni

Jan 6

Everyone is distracted by chatbots. I’m over here orchestrating a 134-modality AGI core.

#agi #python #deeplearning #opensource

2 min read