My journey with AI

OpenAI API beta (pre-ChatGPT)

I came across this email from OpenAI when I was searching for the ChatGPT welcome email. This was more than 1 year before ChatGPT was released.

I was surprised to see that they had a product named Codex back then. This was confirmed by Archive.org capture of openai.com

I have no idea what I did with the API access. I have a faint memory of playing around in the API playground page, but do not remember any details.

AWS CodeWhisperer preview

Tried AWS CodeWhisperer. Posted a demo video on team WhatsApp group showing the tool generating code based on comments.

Also had a faint memory about trying GitHub copilot. But I am not sure and it might actually be the CodeWhisperer usage. Also, remember a few community backlashes against GitHub using code for training around the time.

Tried ChatGPT 1 Month after the launch

First chat: I was working with GraphQL at that time, so tested ChatGPT with a GraphQL question.

Tried Anthropic Claude

Tracked this from the Welcome email from Anthropic.

Cursor - First coding IDE

Tried Cursor for the first time to avoid back and forth copy pastes with ChatGPT

Codeium Windsurf

Loved Windsurf on the first try. Cancelled Cursor, and moved to Windsurf Pro.

Used Windsurf + Claude Sonnet 3.5 to build LinkHQ.co - the startup I co-founded. Most CRUD operations were agentic coded.

Used https://gitingest.com/ to provide context to models when using a less popular library (e.g. Indian payment gateway SDKs).

Role at Mercor as a contractor

Worked as a contractor of a top AI Lab: Got access to an unreleased cloud coding agent from that AI lab. Was part of the initial team that created challenging coding problems that need to be unsolvable by the model (for RLHF). Cannot disclose the lab and coding agent name due to NDA.

Installed Ollama and LM Studio

Tested local models. Remember testing qwen-2.5-coder.

Agno

Built simple agents using Agno. Explored using local models too (via Ollama's OpenAI compatibile API), but could not get tool use to work with local models.

Junie - JetBrains coding agent

Tried Junie on PyCharm and wrote about ithere

xAI Human Data team

Joined xAI's human data team (coding RLHF). Worked with pre-release versions of Grok and evaluated model performance by giving it a range of coding tasks in Python. Used many different coding IDEs, CLIs, VSCode extensions

Built sample MCP servers

Built them to test coding model capabilities while at xAI.

Google Antigravity IDE and Gemini 3 Pro

Tried Google Antigravity IDE and Gemini 3 Pro. Loved the plan mode in the IDE.

LLMs became surprisingly good at frontend

Noticed this while using Opus 4.5 and the frontend design skill. This was a leap I thought would take much longer.

Started using Grok CLI / Grok Build

Internal testing. Got access to Grok Build while working at xAI - evaluated the harness performance, and tested for errors. The tool was released to public on May 2026.

OpenClaw

Installed late Jan, when it was named Moltbot, on an Ubuntu VPS. Did not trust it with my personal accounts. It got a separate email account, a GitHub, and was installed on a remote Ubuntu VPS.

First $50 usage per day.

Created my own OpenClaw Skill

Created "python" openclaw skill:

openclaw skills install python

This was a SKILL.md file packaged for Clawhub with Guidelines to use when writing Python code

First $50 API usage

With OpenClaw started getting $50+ API usage each day on Vercel AI Gateway

Claude Max plan

Moved from Claude Pro to Max - thanks to OpenClaw

Tried Paperclip agent orchestration

paperclipai/paperclip - It was an interesting idea - like Jira, but for Agents. It was satisfying to watch agents working on tasks we assign realtime

Hermes agent

Installed on VPS side by side with OpenClaw.

ChatGPT Pro and Codex migration

Migrated to Codex from Claude due to subscription usage changes.