Vireshwar Das.

Notes from the seams is a small publication edited from Kolkata, Mumbai, and Amarkantak. It has one editor and no staff. It publishes research, short opinion, cultural observation, and tech experiments, and it labels every piece with its provenance so that you can tell what was written by hand, what was produced through an AI-assisted pipeline, and what is an actual working demo.

I am the editor. I go by Vira. Most people who end up on this page are trying to work out whether they should take me seriously. What follows is intended to help.

What I do for a living.

I am currently a Domain Expert at xAI, where I conduct adversarial red-teaming and computer-vision evaluation for the Grok family of frontier models. This is the sort of work that requires telling a very confident machine, patiently and repeatedly, that it is very wrong. I do it with the vocabulary of a working colorist rather than the vocabulary of an ML researcher, and that vocabulary turns out to be surprisingly useful.

I am also the founder of Shot by Vira, a production studio I run out of Kolkata and Mumbai. We handle commercial photography, post-production, and creative-operations retainers for brands who prefer finish over volume. It is a small shop. That is deliberate.

How I got here.

Before xAI, I was Head of Post-Production at Papercloud Films and Studio 1965, managing a sixteen-person team on Netflix, Amazon Prime, and Hulu projects. I cut revision cycles by thirty-five percent through a protocol I wrote and enforced. I mention this because it is the clearest precursor to the work I do now. A frame that needs another revision cycle and a frame that needs another pass through the model are, in the end, the same problem.

Before that, I was Campaign Director at ICE Media Labs, running the ITC Aashirvaad Durgotinashinir Shondhane campaign, which was a large cultural program tied to the Durga Puja season in Kolkata. I ran it end-to-end through a team that was more interested in siding with the client than supporting the work. It shipped on time. Twenty other campaigns followed.

Before that, I was a Key Account Manager at Carl Zeiss India, selling consumer optics across five product verticals and six states. I was twenty percent above target for the year. More importantly, I spent two years learning the grammar of precision optics, which turns out to be the same grammar a computer-vision model is trying, and failing, to speak. The Carl Zeiss years are more directly relevant to my current work than most of my formal credentials.

Before that, I founded the Video Design wing at TresVista Financial Services, produced two hundred animated assets, and wrote the playbook for how the function would work after I left.

What I think the work is about.

Most people agree that art and engineering are opposites. Most people, with great respect, are wrong. I have spent the better part of a decade insisting otherwise, and this publication is in part the argument continuing at a slightly larger scale. Both art and engineering are practices of applying a standard to the world repeatedly, over thousands of hours, until the hand knows the answer before the mind does. The vocabulary differs. The practice does not.

This matters right now, specifically, because frontier AI development is in the process of rediscovering the value of domain expertise after years of assuming it was a rounding error. The people who will build the next generation of useful visual AI are not, mostly, the people currently being hired to build it. I am writing toward that moment.

Where I work from.

I was born in Kolkata and find it difficult to breathe there for long, which is perhaps the most Bengali sentence it is possible to write. I do my best work in Mumbai, which is honest and unsentimental about what it wants from a person. I go to Amarkantak, in Madhya Pradesh, a few times a year. The Narmada begins there. It does not announce itself. I find this consoling and instructive in roughly equal measure.

I read in Bengali, English, Hindi, Marathi, Sanskrit, and clumsy Japanese. I mention this because translation is where a great deal of my thinking about perception actually lives. The gap between how two languages name the same color is, to me, the clearest available demonstration that meaning is made, not found.

How this publication is put together.

The whole operation runs on a pipeline I built with n8n, Claude API, Google NotebookLM, and Google AI Studio. Research notebooks are built natively in NotebookLM. Culture and politics pieces are drafted through a Claude-powered pipeline with a fixed voice prompt, then hand-edited before publication. Life pieces are written by hand, no pipeline. Tech trials are small apps I build on Google AI Studio and Vercel, embedded as live iframes.

Every piece carries a provenance badge. Human-written means I wrote it from scratch. AI-Assisted means the pipeline helped draft it but I reviewed and edited it by hand before publication. Research Digest means it is a summary of a NotebookLM inquiry. Tech Trial means it is a live demo you can run. I want this transparency to be a feature of reading the site, not an apology for how it is made.

If you would like to write.

One address. One person. I answer everything that reaches me. Tell me what you are building, making, reading, or wondering about. I will write back, sometimes briefly, sometimes at length, always in the same voice you just read.

vireshwardas@outlook.com

ভালো থেকো।