31 October 2026 • 7 min read
productgrowth.in is a set of pages for people who build products in India: about 1,000 pages of tools, insights, resources, numbers, prompts and news. It started as about 800 static pages, and those pages are still the base. One person, Rishabh, runs it. The writing, the code and the daily publishing are done with AI tools that he directs.
The pages have not grown on their own. In the week of 24 to 30 October the site had 713 weekly visitors after removing automated-looking traffic, up 26% on the week before. The old static pages were flat (474 clean visitors against 484). The growth came from two newer sections: the news pages went from 56 to 133 clean visitors, and the new numbers pages from 1 to 56. The static pages did not shrink, and they did not grow either.
Every morning a set of timed jobs runs with no one pressing a button. The brief has nine stories by 09:15. The podcast episode is ready by 09:35. A pack of posts, covers and a carousel is assembled by 10:00. The Hindi edition arrives by 11:00. The brief is organised into 12 industries. It has run on 7 of 7 days in the audit week of 21 to 27 October.
The videos are cut from the episode audio. On 29 October all nine vertical videos were finished by 09:45, on one DigitalOcean server with 1 vCPU and 2 GB of memory that costs $12 a month. Our notes say the 1 GB plans are too small for the speech-to-text and video tools together.
The show has two AI hosts, Arjun and Meera. About 160 episodes exist. The podcast page gets about 28 people a month. The videos cut from the show are the only thing that has reached about 70,000 people, which is why the show exists at all: it feeds the videos.
The English episode is voiced once, on an ElevenLabs plan of $20 a month. Hindi and the read-aloud on article pages use Sarvam. The read-aloud audio is made only when a reader presses play, and it is stored, so the second listener costs nothing. We do not voice pages ahead of time.
The ElevenLabs plan has 150,000 characters a month. A daily episode is about 7,000 characters, so a month of episodes would be about 210,000: more than the plan. A guard drops the episode to Sarvam when the ElevenLabs credits run low.
The machine has failed in ways that are easy to describe, and we keep a log of them.
On 22 October both of the 09:15 script attempts failed. One attempt failed on an Apple story that was flagged "never covered" because every word in its title was a stop word. Another failed because the check read "web3" as an unsourced number, "3". The scripts were also 5,879 and 6,170 characters against a 7,200 minimum. The fallback read the headlines aloud, and that version was voiced and published before Rishabh pulled it. The Hindi edition had already been translated from it. The fix that day was to write the script by hand from the nine stories, and to rebuild the videos.
On 23 October we fixed a second problem: the English voice drifted to a different accent partway through. The ElevenLabs dialogue chunks were 2,500 characters with no language set, and their limit is 2,000. They are now 1,800 or fewer, the language is fixed to English, and each episode uses one seed.
An audit of 21 to 27 October found more. The email order did not match the script and video order on 6 of 7 days. Slack showed a tick when the manifest check had failed (26 October). The Hindi edition dropped clips on 4 of 7 days without telling anyone. A spoken number could appear without an alert. A broken (404) picture reached the email, the page, the videos and the cover on 27 October, and again on 28 October. The 404 picture was being fixed on 28 October.
The first talking face we tried used Simli, a service that streams a video avatar. We measured it: 9.5 seconds from the end of a question to the first sound, against 4.9 seconds for our own voice-only build, and 1.2 seconds for OpenAI's real-time voice. We parked Simli. The test showed it was never faster than what we already had.
So we built our own. The face file, face.js, is 35 KB, uses no library and makes no network call. It listens to the sound and moves the mouth, brows, head and eyes; any logo can be turned into a face. A 10-second 720×1280 video took 10.2 seconds to make in a headless browser. It runs in the reader's browser, so it costs nothing per view. It was built on 30 October, and the Intelligence page itself was used by 1 person in the week of 24 to 30 October.
Kriyā is the one product we sell first, at ₹149 a month. It reads our brief, articles and tool pages and answers a builder's question. There is a web app and an Android app. Android version 2.4.3 (build 23) was built and signed on 29 October, with 125 unit tests and a test run on a phone. We have 0 paid users. Signed-out visitors get 3 free answers, which feels like a wall, and of 196 people who saw the sign-up box, 8 signed up.
No single tool does everything. The seats are Claude Code on the Mac, Antigravity for research and drafts, Cursor for the Android app, Claude in Chrome, a Grok agent in Slack that drafts replies, and a Cowork chat connected to Gmail and Drive. Claude Opus plans and checks; Sonnet runs long jobs of several related parts.
The rules for these jobs are written down in one file, and they are short. A job edits only the files it is given. It never commits to the main branch, never pushes and never deploys. A new test must show that it fails when the rule is broken on purpose. Large files are searched, not read whole. Each job ends with a result card: the number it should move, its value before, and the date to look again. Anything that moved nothing is cut or rewritten.
The cost of Claude for all the writing was about $32 in 30 days.
We will not call this a result. A site with 713 weekly visitors and no revenue is a small business that has not started. What the build has shown is narrower: one person can keep a daily machine running with AI, so long as every step has a check, and every failure is written down. The checks catch the AI when it drifts. They do not catch everything, and the log above is proof.
One person, Rishabh, runs it. The code, pages and daily publishing are written with AI tools he directs, and the rules for those tools are kept in one file.
Three costs we track: the server is $12 a month, the English voice is a $20 plan, and Claude's writing cost about $32 over 30 days. We have not added them into one total here. Figures are from our own notes on 31 October 2026.
No. Revenue is ₹0. The target is ₹5,000 a month by 31 December.
On 22 October the fallback headlines-only episode was voiced and published before it was pulled, and the Hindi edition had been translated from it.
Yes: the podcast page, the daily news and the face maker are public.
Nine stories, every morning at 09:15 IST.
Get the brief free →