Déjà vu from 2017, Agentic harness on steroids & Roman aqueducts

Looking back at the blog post I wrote 8 years ago, it is a bittersweet realisation of how certain things never change. A few things improved while others deteriorated.

It has been a long time since I posted anything here. But unlike the last time, I have nothing queued up in my drafts. In 2017, I moved to Bengaluru in search of better career opportunities, and last year I moved to Spain in search of a better quality of life.

And as then, I have been busy trying out a variety of things. Back then I was tinkering with Natural Language Processing among a few other things and now it is with Large Language Models.

And as then, I feel confident of things that I want to keep doing in the future and this is me compiling them here.

Duct - From Supervised NLP to Agentic Harnesses on Steroids

Adam - Stumbling and fumbling with Supervised NLP

During the period of 2017, I was building a smart question-answering system using basic natural language processing and web scraping (limited to Wikipedia). Getting it to answer factual questions seemed like an impossible task, let alone expecting it to do any kind of reasoning. My entire approach relied on using natural language processing to get the system to understand the intent of the question by having it learn grammatical relationships and word embeddings through supervised learning. The Adam project was archived because I did not invest enough in building an active contributor community to take the project forward and shape it up with advancements in NLP.

Around one year into ChatGPT’s first release, I could see the potential of LLMs changing software engineering as we know it. It was more wishful thinking of a lazy individual than some divine vision. And thus we have the next moonshot project.

Tonic - If you could build anything… you could also build it wrong

A modern IDE built from the ground up for LLMs. It was capable of directly intercepting the LSP (a standard that normalises communication between code editors and language analysis tools). It was supposed to have drag-n-drop UI designers, plug-n-play modules, coding assists and all the features that you might find in Cursor IDE. Tonic failed and was abandoned halfway through because it was built on a wrong foundation. I took a huge bet on Flutter (a cross-platform framework to build apps). A much better execution of this can be seen with Cursor (built on top of VS Code).

Duct - Yes, the Roman aqueducts

Not a moonshot project but a simple, opinionated collection of agentic workflows. Workflows I saw myself using often in order to do more with less. And also a sadistic curiosity to push the LLMs to be more and more autonomous and high agency. If you go through my past few months of conversations with Claude Code, you will come across me whipping it with all due respect to be more and more autonomous. Still, every dog has its day.

Duct initially started as a private project to automate some of the organic growth operations, primarily because I sucked at and had a disdain for them. I leave it to you to guess which came first. It expanded from workflows that diagnose SEO issues and identify hidden SEO opportunities to grow, to actually fixing and executing the diagnostics and growth strategies. It later expanded towards content generation for TikTok, and anything and everything that fed my lazy, sadistic desire to make my agents autonomous. It might have something to do with the endless delays in the GTA 6 release or the greed simulator of The Sims 4. I would like to call it open-world simulator withdrawal syndrome.

LLMs are amazing at breaking down and tackling complex problems over the course of time if they are provided with sufficient autonomy to access and gather information and verify hypotheses with well-defined test scenarios (in an isolated sandbox). Bless the wise old sages who preached centuries ago the gospel of Test-Driven Development.

The more robust quality-control systems you put in place, the more robust and well-rounded the solution you will end up with.

I open-sourced Duct knowing that it would be so much fun to build Duct in a community. I would want to use it as an opportunity to not repeat the mistakes from the past with Adam or with Tonic. Check out Duct and give it a star! ⭐

I intend to talk about Duct more actively on the YouTube channel and this blog and a few other places. I am exceptionally proud of how I have set up the project to allow agentic development and I am excited to talk about it soon.

Oh, we did not talk about Roman aqueducts anywhere. You should check them out. They are pretty pretttyy cool!

Other stuff - diving, hiking and more…

I fell in love with diving the day they tea-bagged me in the Sri Lankan ocean (a fun discover dive where you tandem dive at maximum 5 to 10 m depth) and now I am an advanced diver with over 75 dives. I would like to get my cave diving certification done very soon.

I have done lots of hikes in India, and a knee injury put a pause on it, but I have sufficiently recovered. I am back to hiking, and doing more of it. Find me on AllTrails.