Articles Working with Claude IR Scraper Company Deep Dive Music Transcription Sub-Agent Orchestration Hallucinations Two Words for 2024 Dilly-DALL-E-ing

My Experiments with AI

This is a picture from the summer of 2007. When I was half my current age. The image capture the vibe I inhabit these days as I go down many rabbit holes with Claude (Code / Cowork).

IIT Madras CS department, summer 2007 — friends posing for the camera IIT Madras CS department, summer 2007 — the computer lab after an all-nighter

Theoretical Computer Science lab, IIT Madras, ~5 a.m., summer 2007.

These photos above are from the Theoretical Computer Science lab at IIT Madras, which had the swankiest desktop our CS department could afford (a Mac, with a built-in web camera). Probably around 4:30 or 5 a.m., after a bunch of my batch mates and I pulled an all-nighter. Those were the final few weeks of our undergrad days, as we were wrapping up our individual B.Tech projects. Ah, the shenanigans! This was our attempt at a Friends’ style picture.

I find myself in a remarkably similar headspace now — nearly two decades later — with Claude and my AI experiments. The same rabbit hole. You code something, watch it come to life, spot the edge cases, go back to fix it, introduce new errors, rinse and repeat. But the joy of seeing something come alive is real. Like it was, when we wrote code and took part in programming contests.

I just came across this tweet today. Truly feels like it.

Naval on X: Vibe coding is more addictive than any video game ever made (if you know what you want to build).

This will hopefully be an evolving webpage where I want to capture my own learnings as I build these products — a personal log. Something born out of vibe coding itself.

Your first workout will be bad.
Your first podcast will be bad.
Your first speech will be bad.
Your first video will be bad.
Your first ANYTHING will be bad.
But you can't make your 100th without making your first.

— Alec Zamora

Vibe away ;)

Experiments
Working with Claude — A Toolkit After Two Months

The toolkit I’ve ended up with: a working folder, a CLAUDE.md memory file, skills for the work I repeat, voice DNA for the writing. The surprise wasn’t what it taught Claude — it was what it taught me about how I think.
Building a Scraper That Works on Any IR Website

A single scraper for any company's investor relations site. Eight platform types, a Linnaeus-inspired taxonomy, and Taleb's barbell for bot evasion.
Building a Company Deep Dive (Initiation Style) Report

A five-stage workflow for closed-book fundamental analysis — sourced entirely from documents I provide, audited deterministically, no hallucinated numbers.
Transcribing Eight Years of Music Lessons

Using Claude to isolate my teacher's voice from mine in Indian classical music lessons. Three versions, a Tkinter applet, and ground truth validation.
Report Generation and (Sub)Agent Orchestration

Why LLMs are good at generating analysis but bad at assembling it — and how 300 lines of deterministic Python fixed it.
How Bad Are the Hallucinations?

Testing ChatGPT's accuracy on factual questions and understanding when models confidently state false information.

Substack · May 2023
Two Words for 2024

Predictions for how AI would evolve in 2024 — two words that captured the direction of the technology and the market.

Substack · Dec 2023
Dilly-DALL-E-ing

Exploring DALL-E's capabilities and quirks — what works, what breaks, and the gap between possibility and usefulness.

Medium

Want to see more writing? Subscribe to my Substack for regular essays.

Back to Home Subscribe to Substack