Forem: Ayush Gairola

We built our own agent orchestration system!

Ayush Gairola — Fri, 06 Mar 2026 15:36:03 +0000

Straight to the point

I am currently in middle of building AntinodeAI
what is it? Well it is an agentic space for you and your team to leverage the power of AI to research and analyze data at one station.

But I can use other tools, what makes you unique?

Well technically you can use other tools and I will neve lie many of them are very good.
But, Our USP is not AI or Singularity **! but Transparency and performance How? (mentioned below)

Our USP what makes use unique!

Agentic AI is the new frontier of AI that lets the LLMs decide on their own what to do.

To achieve this level of autonomous systems you need complex architecture and systems to handle edge cases, to handle these many frameworks come in play. They let you build these sophisticated systems with ease.
But there is something nobody tells you! It is slow, Yes I said it (roast me haha!).
I chose a different approach why not build my own system, its not that I cannot if I can think of it I can build it and I did In Pure Typescript and JavaScript it allowed me to keep it lightweight, modular, easy to debug, easy to handle LLM behavior and performance.

I also added an event driven system with the help of web sockets that enables ** you to check **what is actually happening behind the scene not just the LLM thoughts but actual process logs, you can toggle it too, so now you know whether the LLM is making things up or actually saying the right things.

What I was able to achieve

We are live in early access phase, with few features running live, web research mode (basic version because of resource constraints not able to run autonomous mode), Synthesis mode , Document based chats and analysis .
Workspace that allows you to access all these features with your team, group and friends.

The transparency layer is the way we want to connect with our users cause nobody wants to see a magical response when they are working on something important they want verification and truth which is so hard to get these days.

What you can expect

You can expect an MVP built by a solo engineer(looking for work), who love building stuff, few bugs (mostly LLM inference rate limits related issues), a fully functional product built with performance in consideration for free with quota due to obvious reasons.

Every feedback will be appreciated and will be major factor in the future growth of AntinodeAI as the resources become slightly more accessible I will be able to focus more on the improvements .
Thanks you!

Why I built an "Anti-Node" for the Deep Web

Ayush Gairola — Sat, 10 Jan 2026 07:24:40 +0000

When I started building AntiNodeAI, I got hit with a reality check: users thought it was just another ChatGPT wrapper.

It wasn't. But my "Developer POV" was blinding me to how people actually perceived the tool. Here is why I built it, the technical hurdles, and why I’m obsessed with the "Anti-Node" philosophy.

The Problem: AI Cheerleaders & Resource Hell

I wanted to learn things from the ground up basics to mastery. But the workflow was a nightmare:

Visit 100+ sites.
Cross-verify conflicting data.
Filter out the SEO fluff.

I tried LLMs, but they kept trying to be my friend.

"I HAVE BEEN THERE!"
"I AM GLAD YOU ASKED THAT!"

I don't want a personality; I want a research engine. I needed something that stayed in context, didn't hallucinate "friendship," and focused on speed.

Why "AntiNode"?

The name comes from physics. In a standing wave, the Antinode is the point of maximum amplitude the point of maximum energy and vibration.

A node is a point along a standing wave where the wave has minimum amplitude. For instance, in a vibrating guitar string, the ends of the string are nodes. By changing the position of the end node through frets, the guitarist changes the effective length of the vibrating string and thereby the note played. The opposite of a node is an antinode, a point where the amplitude of the standing wave is at maximum. These occur midway between the nodes.

View on Wikipedia>

I wanted to build the point where information is at its highest "amplitude" where the noise is filtered out and the signal is strongest.

The Tech Stack & Pipeline

1. Human Verified RAG

I started with a RAG system where you upload docs and select privacy modes. But I hit a snag: The "Garbage In, Garbage Out" problem. If a user uploads junk, the UX dies. So I added a Human Verification Layer. Now, responses show the specific sources used, complete with upvotes and downvotes. It’s a decentralized trust layer for AI.

2. Deep Web & Synthesis Mode

Waiting for manual contributions is slow. I added a deep-web search feature using a custom scraping pipeline that prioritizes privacy.

The coolest part? Synthesis Mode. Summarizing one doc is easy. Cross referencing five different file formats at once is hard. I built an NLP pipeline where an agent handles the context switching for you:

# Example Prompt
"Generate a comparative report on doc1.txt and doc2.md vs current web trends"

The agent calls the necessary tools, manages the context window and information extraction, and synthesizes a single, factual response.

Shipping > Polishing

The features are live. Is it buggy? Probably. I focused on shipping the core logic rather than making the buttons pretty.

I’ve kept third-party libraries to a minimum. I’m currently running this on a low-end setup, but my goal is to eventually move toward a fully custom built stack. Why? Because I like knowing exactly how my code handles every bit.

Check out the progress:

https://antinodeai.space

What I’ve learned so far:

Context > Personality: Stop making AI "nice" and start making it useful.
Collaborative Research: I added "Rooms" because great things aren't achieved in a vacuum.
The "Wrapper" Trap: If you don't show the "how" (sources, verification, custom pipelines), people will assume you're just hitting an API endpoint.

Have you ever struggled with "selling" the technical depth of your project? How did you bridge the gap?

Beyond the Code: What I Learned Building My First SaaS

Ayush Gairola — Thu, 08 Jan 2026 20:03:22 +0000

The "Eureka" Moment

When I first started learning to code, I genuinely couldn't understand how people enjoyed it. It felt like a chore. But then, I built my first project. After wrestling with the basics, I hit that first Eureka moment. I realized: I actually enjoy this.

The "I Can Build Everything" Phase

As I kept learning, I’ll admit—I got a bit cocky. I thought I could build anything from scratch. I convinced myself I didn't need libraries or APIs; I was going to do everything myself.
Inevitably, I landed in a pit of my own making. That was my first major lesson: It is not always necessary to reinvent the wheel.

Solving My Own Problem

I decided to work on an actual SaaS based on a struggle I faced personally. I had two main pain points:

Resource Overload: How hard it is to find resources, cross-check them, sort them, and verify them before even starting to learn.

LLM Personality Bloat: I use LLMs for help, but I can't stand it when they try to act human. I don’t need a machine saying, "Great question! I’m so glad you asked that!" I want a research tool, not a cheerleader. I want it to stick to the context, not jump between personalities.

The Reality Check

I built the tool, started using it, and it was exactly what I wanted. But when I asked others for a review, they were lost.

I had made a classic mistake: I was selling a service from a developer's point of view. I wasn't speaking to the non-technical users who didn't understand how the tool actually solved their problems.

The Final Realization##
Coding a product and actually selling a product are two entirely distinct tasks. Both require specialized skills and immense dedication. Neither is "easier" than the other.