I build AI agent products end to end at Gradscaler. I got here the long way — WordPress, then Flutter, then backends, then agents — and freelancing taught me the other half of the job: turning what a client is actually asking for into something that ships.
A few things I've built — client work and my own. The client repos are private, so these describe the work rather than link to it.
Separating parents each talk things through alone with the agent, in their own private room. Once both are done it lines the answers up — agreed, partly, not yet — and hands that to their mediator for the real conversation.
Live for Better Parenting Plan. The first version ran the joint session too; mediators said that was a step too far, so I rebuilt it around them. Thirteen topics, two scoped rooms, LangGraph and FastAPI underneath.
in productionYour salary lands, you confirm it, and it splits by percentage into buckets — savings and investments first, expenses get what's left. No bank linking, nothing logged behind your back.
It's the Notion tracker I ran on myself for two years, turned into a real app. A hundi is a medieval Indian bill of exchange, and also a temple collection box — which is where the piggy banks come from.
sole author · in betaA voice agent that lives on my Linux home server and never phones home — named after the assistant in Iron Man. Ask it about the machine and it answers out loud while the terminal fills in the detail.
The fun part is the budget: 4 GB of VRAM. Ollama takes the GPU, Whisper and Piper stay on CPU so they're never fighting over it. Read-only by design — it can tell you about the server, never change it.
fits in 4 GB VRAMGives Claude a video editor. Describe what you want and it scaffolds the Remotion project, writes the scenes, syncs narration to the words and renders the MP4. Thirteen tools, one composition.json holding the truth, no timeline to drag.
Every step in the timeline below started as something I taught myself on the side, and the working out is public. These are the repos where I actually learn things — commits from the days I hadn't understood it yet, not a tidy write-up after the fact. When a colleague picks up the same tool, this is what I point them at.
Agent graphs, from the first stateful node to the patterns that ended up in production.
Where the agent work started, a month before LangGraph and two before the job title.
The current one — working through Agno to find out what it does better than what I already use.
Notebooks from going back to the maths underneath all of it, rather than staying at the library layer.
Mechanical engineering degree, WordPress first job, AI agents now. Nothing here was a straight line — every step started as something I taught myself on the side.
I was studying mechanical engineering and building sites on the side — WordPress and Figma for ArctikCircle, plus whatever I could teach myself at night. The handle on this site is left over from that first team.
Still writing WordPress for a living, I spent evenings drawing Sierpinski triangles and Barnsley ferns with Flutter's CustomPainter, purely to learn it. The commits are still public. Seven months later that was the job.
Mobile developer first, Flutter and Kotlin. Then at Alendei I moved off the web team into support and built a customer lifecycle system on Zoho webhooks and the WhatsApp API. First time I owned the whole path — request in, data out, and a real person on the other end waiting on a reply.
Same move as Flutter, second time around. I started pushing agent experiments to my own repos in September, LangGraph a month later, and by November it was my actual title at Gradscaler. Learning in the open, slightly ahead of the job, is the only career strategy that has worked for me twice.
Freelance work put me in front of people who know their business inside out and know nothing about software. Learning to sit in that conversation and come back with something buildable turned out to matter as much as anything technical — it's the reason the Gabby rebuild happened at all, because I could hear what the mediators were really asking for underneath the feature request.
Gabby for Better Parenting Plan is the clearest example — I wrote most of it, across a seven-person codebase, then built the follow-up when real mediators told us the AI was doing too much. These days the work runs from the first conversation with a client to the thing that actually runs in front of their users.
What I reach for by default. The list is short on purpose.
Have something that needs building and only a rough idea of the shape? That conversation is the part I like most. Tell me what the problem is and I'll tell you honestly whether I'm the right person for it.
Based in Vadodara, India — UTC+5:30.