
Report First, Figure It Out Later: How I Taught AWS Lambda to Be a First Responder 🚨
Turns out the hardest part of "just report the problem" was never the reporting. It was the deciding.
It's 9:47 PM in a hostel. Someone notices the ceiling fan in Room 204 has started making a noise like it personally disagrees with the concept of ceiling fans. Nobody knows who to tell. Is this a "text the warden" situation? A "call maintenance whenever, no rush" situation? Or a "this thing is about to become a projectile" situation?
Nobody's sure, so naturally, the correct hostel response kicks in: everybody ignores it. For three more days. Until the fan eventually resolves the debate itself, dramatically, at 2 AM, and suddenly nine people are very interested in who should have been told.
This is not a fan problem. This is a decision problem. Every "report an issue" system I've ever met has the same personality flaw: it hands you, the panicking human, a form and says "please correctly diagnose your own emergency before we help you," which is a bit like a fire extinguisher asking you to first fill out a risk assessment.
So for Bharat Builds, I built something that just listens instead. You tell it what's wrong. It figures out the rest. No dropdown stands between you and getting help, mainly because dropdowns have never helped anyone in a crisis, ever, in the history of dropdowns.
"The hardest part of reporting a problem was never describing it. It was knowing who's supposed to care."

1. The Great Lie of Every Report Form Ever Made
Every form based system quietly assumes you, a stressed out human mid-situation, are also somehow a calm, well-informed civil servant who knows:
- exactly how serious this is (you don't, you're mildly panicking)
- exactly which department owns it (you have never once known this)
- and has the patience to scroll a dropdown menu labelled "Category of Concern" while something is actively on fire, occasionally literally
A leaking tap, a suspicious noise, and an actual fight near the gate are not the same emergency, and yet most systems make you sort that out yourself before they'll even pretend to help. That felt backwards enough that I decided to just remove the sorting entirely. The pitch became almost insultingly simple:
"All I need to do is report what is happening."
No category. No severity slider. No "which team should handle this," a question I'm fairly sure even the teams themselves can't answer half the time. Just: what happened. The AI reads the vibe. The system takes it from there.

2. Understand, Assess, Decide: Basically a Very Judgmental Triage Nurse
Imagine walking into a hospital and just saying "my chest hurts," instead of first self-diagnosing into a medical category you definitely made up. A good triage nurse doesn't hand you a 40 item checklist. She looks at you for four seconds, asks one sharp question, and decides: sit and wait, fast track, or straight into the scary room.
That's the whole engine, dressed up in AWS instead of a stethoscope. And crucially, it's not one all powerful brain making the call. It's three separate, mildly suspicious of each other components:
- The AI listens to your chaotic, possibly typo ridden description and turns it into something structured: what happened, how bad, how sure it is.
- A boring, rule following Decision Engine, containing zero AI and zero opinions, looks at that and decides the actual path: sort yourself out, keep an eye on it, or wake somebody up.
- An actual human is the only one allowed to push the big red button.
"AI interprets. Software decides. Humans authorize sensitive actions."
This sentence is basically the constitution of the whole project. The AI can be as clever as it wants about understanding "there's a fight near the main gate, someone might be hurt." What it does not get to do is then also decide to dispatch someone, because I would like to sleep at night, and also because that's how you end up on the news.

3. Choose Your Own Adventure: Three Endings
🟢 Low severity, sort it yourself. "The tap in Classroom 204 is leaking slightly." Congratulations, nobody needs to be summoned for this. You get calm instructions and a gentle mention of whose job it technically is, so you can feel smug later.
🟡 Medium, keep watching. "That fan noise, again, day three." Not quite a crisis, more of a slow burning subplot. The system remembers it happened before, connects the dots, and quietly flags the right team before it graduates into a full incident.
🔴 High or critical, wake somebody up. "Active fight near the main gate, someone may be hurt." Here the system drops the friendly chatbot act completely, asks for your location like it means business, figures out who's actually responsible, and stages a handoff that a human still has to approve. It does not get to be a hero on its own.
That last lane is the whole reason this project exists.

4. The Boring Table Nobody Asked For, But Judges Secretly Love
I'll keep this part short because tables are not funny, they are just tables.
AWS Service | Job | Why it's here and not decoration |
|---|---|---|
API Gateway | Front door | Keeps the internet from directly touching anything important |
Lambda (Ingestion) | Writes the report | Deliberately dumb and fast, no thinking allowed here |
DynamoDB | Source of truth | The one place not allowed to shrug and say "probably" |
DynamoDB Streams | Notices changes | Turns "something happened" into an actual signal |
EventBridge | Office gossip network | Lets the AI, search, and review workflow ignore each other's existence |
SQS + DLQ | Waiting room | AI thinking is slow, so nobody submitting a report should have to watch a spinner |
Lambda (AI Worker) | The actual brain | Sits behind the queue so it can fail quietly instead of loudly |
Step Functions | Patient babysitter | Human approval can take a while, and this is fine waiting |
OpenSearch Serverless | Long memory | Remembers "wait, didn't this exact fan complain last week too" |
S3 | Evidence locker | Photos and audio go straight in, nobody drags them through Lambda |
If I had to defend one choice in front of a judge with crossed arms, it's SQS sitting between ingestion and the AI worker. Nobody reporting an emergency should be staring at a loading spinner while a language model thinks deep thoughts. The write happens, the user is free, the AI catches up whenever it's done being clever.

5. The Part That Kept Me Up At Night, Genuinely
There's a big difference between an AI that's allowed to suggest something and one that's allowed to just go do it, and only one of those should ever get near an actual emergency. Guess which one I picked.
The AI never touches the decision directly. It's only allowed to output rigid structured JSON, severity, confidence, a plain summary, and that JSON gets interrogated by schema validation before the Decision Engine even glances at it. Fail the schema, get zero votes. The routing logic itself is proudly, aggressively boring, hardcoded rules with no personality, exactly what you want deciding whether someone gets woken up at 2 AM.
High risk cases don't get a friendly push notification and call it a day. They trigger a Step Functions execution that just sits there, patiently, sometimes for ages, until an actual human clicks authorize. The AI can build the entire case file. It does not get to close it.

6. Lessons Learned, Mostly By Breaking Things
- Doing AI work asynchronously means the frontend has to politely keep asking "are we there yet," instead of getting an instant answer. Worth it, every time, to never block someone hitting submit.
- Adding OpenSearch next to DynamoDB felt like showing off until the exact moment I needed to answer "hasn't this fan complained before," at which point DynamoDB just shrugged and OpenSearch saved the day.
- If I rebuilt this today, the AI worker would get proper dead letter alerting instead of silently dumping failures into a queue I only remembered to check at 1 AM, purely by accident, while debugging something else entirely.

7. Why Any of This Matters
Back to that hostel room and the fan with opinions. Nobody there failed to care. They failed because reporting the problem and figuring out what the problem meant were stuck together as one exhausting job, dumped entirely on whoever noticed it first.
This splits that apart. You describe reality, badly if you must, typos and panic included. The AI understands it. The engine decides what happens next. A human signs off on anything serious. Nobody has to correctly diagnose their own emergency just to get someone, anyone, to listen.
Somewhere, a ceiling fan is still making that exact same noise. At least now, on day one instead of day three, someone actually finds out.

Tech stack: React + TypeScript (frontend). Python on AWS Lambda + API Gateway (backend). DynamoDB + S3 + OpenSearch Serverless (data). EventBridge + SQS + DynamoDB Streams (eventing). Step Functions (human workflow). AWS SAM (infra as code).
Repo: github.com/jackso1328/report-to-me-ai
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article