
CareSetu: Building the Bridge Between Hospital and Home, in 72 Hours
A discharge summary is dense, clinical, and often in the wrong language for the person who has to act on it. Here's how we built an AWS-powered pipeline that turns it into a checklist a patient or caregiver can actually follow - and everything that broke along the way.
The problem nobody talks about at discharge
Everyone talks about the moment a patient gets admitted to a hospital. Almost nobody talks about the moment they leave.
You get handed a discharge summary dense, clinical, often in English regardless of what language you actually think in and you're expected to correctly interpret medication schedules, follow-up dates, warning signs, and dietary restrictions while you're stressed, tired, and focused on just getting home. Miss a dosage timing. Miss a follow-up date. Miss a warning sign that actually mattered. The hospital's job ends at discharge; the patient's job safely acting on everything in that document is just beginning, often without any support at all.
That gap is what my teammate Kamal and I set out to close at First Commit, WeMakeDevs' AWS Bharat Builds Tour hackathon, building CareSetu -> your bridge from hospital to home recovery.
What CareSetu actually does
The core flow is simple to say, harder to build correctly:
Upload → OCR → Simplify → Translate → Checklist → Remind → Support
A patient or caregiver uploads a photo or scan of their discharge summary. Amazon Textract extracts the raw text. Amazon Bedrock rewrites it in plain language. Amazon Translate converts it into the patient's language of choice we support English, Hindi, Punjabi, Kannada, Malayalam, Tamil, and Telugu. The result becomes a structured care plan: medicines, daily tasks, follow-up appointments, warning signs, and diet guidance with automated email reminders ahead of every follow-up date.
One design decision we made early and didn't compromise on: the original document is never discarded. It's always one tap away from the simplified version. And anything the AI is genuinely unsure about -> a dosage that's ambiguous, an instruction that doesn't parse cleanly gets flagged for the patient to confirm with an actual clinician rather than silently guessed at. CareSetu explains and organizes what's already written. It never diagnoses, prescribes, or substitutes a doctor's judgment. That boundary isn't a legal disclaimer bolted on afterward - it shaped how we designed every screen.
The architecture
We split the work along a clean seam: I owned the full-stack application (React frontend, Node/Express backend, REST APIs, the entire patient-facing experience), Kamal owned the AWS/AI pipeline (Textract, Bedrock, Translate, DynamoDB, and the reminder infrastructure). That split let us work in parallel almost the entire hackathon without blocking each other - as long as we both built strictly against a shared API contract from day one.
One architectural call worth mentioning: our frontend allows uploads up to 20MB, which is too large to route directly through a Lambda/API Gateway request body. Instead of forcing every discharge photo through that bottleneck, I built a two-step presigned-URL upload flow — the frontend asks the backend for a signed S3 URL, uploads the file directly to S3, then tells the backend the upload is done so the pipeline can kick off. It adds one extra round trip, but it means the backend never has to touch the raw file bytes at all.
What actually went wrong (the part that matters more than the demo)
Anyone can write a features list. What I think is actually useful to share is what broke, because that's where the real learning happened.
Textract and native-language documents were slower than expected. Our very first real test extracting text from a discharge document in a regional language rather than plain English took noticeably longer than we'd planned for, and it was the thing that ate the most time on day one. It's easy to design a pipeline assuming every step is roughly the same speed; it's another thing to actually run it against messy real-world input and watch where it bottlenecks.
A working feature quietly turned into a hard-coded landmine. Midway through, our "Original Document" viewer started rendering a raw AWS
AccessDenied XML error instead of the actual document. Not a UI bug a permissions bug. The document was being fetched from a raw S3 URL instead of a signed one, which works fine right up until it doesn't. The fix reinforced something worth remembering: anything touching S3 objects needs to go through presigned URLs from day one, not retrofitted after a judge tries to click "view original" and gets an XML error dump instead of a PDF.A blank white screen taught me a lesson about try/catch's limits. We built a map feature using Leaflet, wrapped in a try/catch specifically so a map failure wouldn't crash the whole app. It still crashed the whole app with a blank white screen and no error visible anywhere in the UI. The reason:
leaflet and react-leaflet weren't actually installed yet, just referenced in code. A try/catch only guards against runtime exceptions; it can't catch a bundler failing to resolve a package that was never npm install -ed in the first place. That failure happens one layer below your code, before your code even runs. It's a small distinction that cost real debugging time to actually pin down.The UI initially looked exactly like what it was: AI-generated, fast, and generic. Pastel cards, rounded icon badges, soft contrast the recognizable signature of a first-pass component-library build. That might be fine for an internal tool. It's not fine for a product meant to be read by rural patients and elderly caregivers alongside educated urban users, where font size, contrast, and unmistakable visual signals for "this is important" aren't cosmetic choices they're the whole point of the product. We treated the redesign as seriously as any backend bug fix, not as decoration.
What's next
CareSetu is deliberately scoped for a 3-day build, and we're honest about what's roadmap versus what's shipped. Sessions are currently anonymous and browser-local - no login, by design, to keep the flow frictionless - but that also means clearing your browser loses your care plan, which is a real limitation for a production version. Automated WhatsApp reminders are on the roadmap but blocked on Meta Business API verification we don't have yet; manual prescription-sharing over WhatsApp via a simple click-to-chat link is fully live today. And Find Pharmacy currently uses representative demo data rather than a live geolocation search - a deliberate scope cut, not an oversight, made explicitly so we could spend our limited hackathon hours on the core pipeline instead.
Why this, why now
Healthcare communication gaps aren't a hypothetical problem in India - they're a daily one, and they disproportionately affect the people least equipped to push back on confusing paperwork: elderly patients, first-generation-literate caregivers, anyone managing a relative's recovery without a medical background. CareSetu isn't trying to be a diagnostic tool or a doctor replacement. It's trying to be the thing that stands between a patient going home with a piece of paper full of jargon, and a patient going home with something they can actually follow.
Built by Hardik Khanduja (full-stack/product) and Kamal Kumar (AWS/AI infrastructure) for First Commit - AWS Bharat Builds Tour, organized by WeMakeDevs.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article