AI medicine-awareness bot worked — but only in English, the one language most of the people I built it for can't read. Here's the problem, the one-day sprint to fix it, and how GitHub Copilot got me through it.

GitHub Finish-Up-A-Thon submission. I shipped this project months ago, but it was never truly finished — it was missing the one feature that decided whether a real person in my own community could use it. This challenge was the push that made me sit down and close that gap in a single night.

The problem we face

In India, you can walk into most pharmacies and buy prescription drugs without a prescription. We've quietly normalized something genuinely dangerous, and the research is blunt about it.

A simulated-patient study in Bengaluru sent two researchers into 261 pharmacies with fake symptoms. Antibiotics were handed over without any prescription at roughly two-thirds (66.7%) of them. Not a single pharmacy warned about side effects. Only about one in five even mentioned that a doctor's prescription was needed. The "guidance" most people walked out with was "take it twice a day" — and nothing else. (study)

Here is the part that kept me up at night: it's not because people are careless. The shopkeeper isn't a doctor. The label is printed in tiny English. A proper consultation costs time and money many families simply don't have. So people take what they're handed — and hope it's fine.

Why it's important

This isn't a small inconvenience. Taking the wrong medicine, or the right medicine the wrong way, is one of the biggest drivers of antibiotic resistance — the slow disaster where medicines stop working.

The landmark 2019 Lancet GRAM study estimated 1.27 million deaths worldwide were directly attributable to drug-resistant infections in a single year. In India alone, an estimated 297,000 deaths were directly attributable to AMR, and about 1.04 million were associated with it. (Lancet GRAM 2019, India figures)

I'm not a doctor and I can't change how medicines are sold. But I'm a developer, and I kept circling one small thing I might actually be able to build:

What if a person could at least understand what they're holding — before they take it?

So I built Dawa Saathi. You send a photo of a medicine label, and it reads it with AI and explains, in plain language: what the medicine is commonly used for, its common and serious side effects, and the key warnings (pregnancy, liver, kidney, alcohol, children, the elderly). It does not diagnose. It does not give a dose. It never says a medicine is "safe to take." Because with medicine, a confident wrong answer is more dangerous than no answer at all.

The gap I shipped with (the "before")

Here's the uncomfortable truth about the first version: it only spoke English.

I built Dawa Saathi for people in rural and small-town India — the elderly, people with little formal schooling, people who grew up reading Kannada, Hindi, or Marathi, not English. I come from that world myself; I went to a Marathi-medium government school in a village before I ever wrote a line of code. And yet the tool I built for those people answered every single question in fluent English.

It worked. The AI read labels accurately, the safety guardrails held, caching and rate limits were solid. But it was "finished" the way a bridge is finished when it stops halfway across the river. The one feature that decided whether a real person in my community could actually use it — language — was the thing I'd left undone.

The Finish-Up-A-Thon prompt ("revive a project you started but never completed") landed on exactly that nerve.

One day, under pressure

I gave myself the length of the challenge to fix it, and it came down to a single overnight session — about six hours straight.

That kind of pressure cuts both ways. On one hand, a deadline is the best forcing function I know; the project had sat "good enough" for months precisely because nothing forced me to finish it. On the other hand, language support is not a cosmetic change — it touches the welcome flow, every button, every error message, the medicine response itself, and the storage layer. At 2 a.m., with a medical app where a sloppy mistake isn't just a bug, that's a lot of surface area to get right while tired.

There were stretches where I went down the wrong path — I first started building a translation dictionary, realized it would silently drop medical words, and had to back it out and rethink the whole approach. Hours like that are where solo projects usually die.

How GitHub Copilot helped me through it

I enabled GitHub Copilot specifically for this sprint, and I'll be honest about exactly what it did — good and bad — because that's the point of this challenge.

Where it genuinely sped me up: the repetitive scaffolding. The new i18n layer has dozens of string entries across two languages; the Telegram handler needed every reply rewired to read from it; the callback router needed a whole new branch for language selection. From a comment and the surrounding code, Copilot autocompleted big chunks of that pattern-heavy work. Tasks that would have been an hour of careful typing became a few minutes of accepting and fixing.

Where it helped me think under fatigue: when the bot stubbornly kept replying in English even after I'd changed the code, Copilot Chat helped me reason through the deploy pipeline and spot that I was running an old compiled build (npm start runs dist/, not src/ — I'd forgotten to rebuild). At 3 a.m., that saved me from chasing a ghost for an hour. It also caught a sloppy callback branch where my "language" and "purpose" buttons could collide.

Where I had to stay in charge: Copilot is confidently wrong sometimes. It suggested code that looked right but quietly dropped a safety guardrail, and once invented an API shape that didn't exist. In a normal app that's an annoyance; in a medical app, an unreviewed AI suggestion is exactly the "confident wrong answer" my whole project exists to prevent. So I read every line it gave me. The speed was real — but the judgment stayed mine.

Honestly, Copilot didn't replace that six-hour night. It made the difference between finishing it and giving up around hour three.

The major update: what actually changed

The first thing the bot now does on /start is ask which language you're comfortable in:

🌐 Please choose your language
ದಯವಿಟ್ಟು ನಿಮ್ಮ ಭಾಷೆಯನ್ನು ಆಯ್ಕೆಮಾಡಿ
[ English ]   [ ಕನ್ನಡ ]

Tap a button, and everything afterward — welcome, buttons, questions, and the full medicine explanation — comes back in that language. The choice is remembered, and you can change it anytime with /language.

Under the hood, I split the text into two kinds and treated them very differently:

1. Fixed UI text (buttons, labels, errors, and the all-important medical disclaimer) lives in a small i18n table. It's instant, costs zero tokens, and — critical for health content — the disclaimer in each language can be reviewed word-for-word instead of machine-guessed.

2. The medicine explanation is generated by the LLM directly in the user's language. I deliberately did not translate the English afterward. The awareness model writes its answer in Kannada from the start, while keeping JSON keys in English (so the parser never changes) and drug names in Latin script (so "Tranexamic Acid 500mg" is never mangled). Every safety rule — no dosage, no diagnosis, no "safe to take," always "consult a doctor" — is repeated for the target language.

That choice means one model call instead of two, no brittle dictionary dropping medical words, and a reply that reads like it was written for a Kannada speaker rather than awkwardly converted into Kannada. Each language is cached separately, so the next person who scans the same strip gets an instant answer.

What this means for people in Karnataka

This is the part I actually care about.

Kannada is the first language of around 44 million people, and roughly two-thirds of Karnataka's ~68 million residents speak it natively (2011 census / overview). Meanwhile, by census-based estimates only about one in ten Indians use English at all — and a far smaller share read it comfortably enough to parse a medicine label or a list of side effects.

Put those two facts together and the old version's flaw is stark: an English-only health tool quietly excluded the majority of the exact people it was meant to help. The grandmother in a Belagavi village holding a strip of tablets, the daily-wage worker in Kalaburagi who left school early — they could open the bot, but they couldn't read it.

With this update, that same grandmother can now ask in Kannada and get an answer she can actually understand — what the medicine is for, what to watch out for, and a clear, repeated nudge to see a doctor. It doesn't make the bot a doctor. It makes the bot legible to the people who need it most. For a medicine-awareness tool, legibility in the user's own language isn't a feature — it's the whole job.

The before → after, at a glance

Future plan

• Hindi and Marathi next — the architecture already supports them; each new language is roughly a two-line change plus a reviewed disclaimer.
• Voice output, for users who can't read at all — arguably the people who need this most.
• Reviewed translations of the fixed disclaimer with help from native-speaker volunteers, so the safety text is never just my best guess.
• Pharmacy and NGO partnerships — the long-term home for something like this is alongside health workers and CSR programs, not as a lone bot.

The honest impact

I want to be careful not to overclaim. This won't fix the bigger problem. That needs doctors, pharmacists, and real regulation. One Telegram bot does not move a national health statistic.

But the gap it closes is real and specific: a person can now understand, in their own language, what they're about to put in their body — before they take it. If that makes even a few people pause and think, or take one question to a doctor they'd otherwise have skipped, it was worth the night.

If you work in public health or medicine, I'd genuinely value your guidance on doing this responsibly.

Try it / code

• Bot: TELEGRAM_BOT_LINK — https://t.me/dawasaathi_bot
• Source: https://github.com/SurajKhonde/Dawa_saathi_mediAssist (MIT licensed)

• Demo:

  • Before_challenge:-https://drive.google.com/file/d/1UW5HsSfBZKWx93P-flJ9sUKVrg03JflS/view?usp=sharing
  • After_challenge :- https://drive.google.com/file/d/1pDaeMIelnb9ymHF9UygBmR4DY6lT7Vh1/view?usp=sharing

Built solo. Not a substitute for a doctor — by design.

Sources

• Non-prescription antimicrobial dispensing in Bangalore pharmacies (261-pharmacy simulated-client study): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4676185/
• Global burden of bacterial AMR, 2019 (Lancet, GRAM): https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(21)02724-0/fulltext
• India-specific AMR mortality (297,000 attributable / 1,042,500 associated, 2019): https://journals.lww.com/pmrr/fulltext/2024/01040/antimicrobial_resistance__an_emerging_global.1.aspx
• Kannada speakers and Karnataka share (2011 census overview): https://en.wikipedia.org/wiki/Kannada