Connecting the dots
“You can’t connect the dots looking forward; you can only connect them looking backward.”
— Steve Jobs
Hardi Jain · AI and SaaS product manager
About half of people on long-term medicines don’t take them correctly, often because the label can’t be read. I wanted to see how far a product manager could get with AI tools and no engineer: a prescription photo turned into clear spoken instructions, built in an afternoon and shipped to a live URL.
01 · CONTEXT AND ROLE
I wanted a sharp test: could I take a real, evidenced problem from idea to a deployed, working product with AI tools alone, in about two hours?
I chose the problem, wrote every spec and prompt, reviewed and tested what came back, and shipped it. I wrote no code by hand.
02 · THE PROBLEM
The WHO puts adherence to long-term medication at about 50%. In India, only 40 to 60% of older adults take their medicines correctly, and low health literacy comes up again and again as the reason.
A label is small, clinical, and often printed in a language the patient doesn’t read well. “1 cap PO BID p.c.” means nothing at the kitchen table, and “twice daily” never says when.
03 · THE INSIGHT
Every other medication app assumes you can already read and understand the label, then adds reminders on top. For the person I had in mind — a senior who reads slowly, in Tamil, not English — that assumption is the whole problem.
So the product does one thing well: it reads the label for you and speaks it, in your language, and only then reminds you.
04 · THE SOLUTION
Photograph the label. No typing, no searching for the medicine by name.
It reads and simplifies. AI pulls out the medicine and dose, rewrites the instructions in warm, plain language, and splits the day into meal-based times.
It speaks, in 11 languages. Hindi, Tamil, Bengali, Spanish and more, with a phonetic version behind the scenes so any phone pronounces the script correctly.
It reminds. A voice alarm at each dose time, with the one safety warning that matters pulled to the top.
05 · HOW I BUILT IT (NO-CODE)
I described the app in plain English to Google AI Studio and had a working prototype the same afternoon. The vision AI (Gemini) does the hard part — reading the label, simplifying, translating — so there was no OCR or model to train.
The prototype’s AI logic only ran on my machine and would have failed the moment it deployed. I directed a coding agent to move it into a serverless function so the live app actually works, then shipped it to Vercel.
Every change followed the same routine: I wrote what it had to do, the agent built it, and I tested before moving on.
06 · KEY DECISIONS AND TRADEOFFS
Voice-first, not app-first. A less “appy” interface, in exchange for something a low-vision senior can actually use.
Lead B2B2C, through pharmacies. The person with the problem isn’t the one who pays; pharmacies and caregivers are. That reframed the go-to-market.
No invented market size. I documented the sizing method and inputs to validate instead of a number I couldn’t stand behind.
Flag the safety fallback. When the AI is busy the app shows sample guidance, clearly marked as a sample, never as a real reading.
07 · CHALLENGES I SOLVED
The AI worked here and 404’d live. The label-reading endpoint only ran in local dev. I traced it and moved the logic into a serverless function so it exists in production.
The offline fallback showed the wrong medicine. When the AI was busy it returned a generic sample that a user could mistake for their own label. I flagged it so a sample can never pass as real.
The uploaded source broke the build. The exported files had lost their folder structure. I restored it so a clean build passes.
Untested across languages. I wrote a repeatable test that checks all 11 languages return complete, correctly-scripted, speakable text.
08 · IMPACT AND WHAT I’D MEASURE
There are no usage numbers to report. What exists: a deployable app that reads a label and speaks it in 11 languages, with meal-based reminders and a fallback that’s honest about when it can’t read the label.
Rather than claim results it hasn’t earned, I defined the metric that matters: doses each week where someone got accurate guidance and then took the medicine. Next is a small test with real seniors and caregivers.
09 · REFLECTION
The demo took an afternoon. Making it safe took the real work: what happens when the AI is busy, how to keep a sample from looking like a real reading, how to sound right in a language the phone can’t render, and how to fail honestly.
The lesson repeats from my last build: read what the first version actually does, and test on the real hosting from the first change, not the fifth.
10 · SKILLS
Product: framing a problem with cited data, cutting scope to one job, choosing a metric that tests the core assumption, and picking a business model that matches who actually pays.
Working with AI: writing specs an AI can build from, reviewing output instead of trusting it, and catching a prototype’s fake features.
Enough technical depth to run a build: spotting a dev-only API, moving it to a serverless function, and writing a test that guards all 11 languages.
Judgment: refusing an invented market size, and making a safety fallback honest rather than seamless.
Product Manager · AI and SaaS
Connecting the dots
“You can’t connect the dots looking forward; you can only connect them looking backward.”
— Steve Jobs