
How I built an AI agent that explains scary messages to my parents — in one day, for ₹0
Starting from zero AWS knowledge, I built Saral — a multilingual, voice-enabled AI agent that explains confusing bank SMS, notices, and bills in plain language. Here's exactly how, step by step, running fully local for ₹0.
I'm going to be honest with you about where I started.
A day before this, I had never touched AWS. I had made an account, panicked that money would get deducted, and hadn't even claimed my credits. I didn't have a project idea. I'm a firmware and FPGA person — RTL, Verilog, waveforms — cloud and AI agents were not my world at all.
By the end of the day I had a working, multilingual, voice-enabled AI app running on my own machine, costing me nothing. Here's exactly how, step by step, in case you're starting from the same place I was.
The problem I picked
I didn't want a fake "hackathon idea." I wanted something real. And the realest thing I could think of was my own parents.
They can use a phone fine. But when a bank SMS, a government notice, or a bill shows up — full of English and jargon — they don't know what it means or what to do. Worse, fake "your KYC is incomplete, click this link" scam messages look exactly like the real ones. People lose money to this every single day.
So I built Saral — an assistant where you paste, speak, or photograph a confusing message, and it explains in plain language, in your language, what it means and exactly what to do next. It even reads the answer aloud, for people who can't read well.
The zero-cost architecture
The key realization: I didn't need a paid cloud model. The whole thing runs locally.
User (Streamlit UI) → AWS Strands Agents SDK → Ollama (local LLM, my own machine)- AWS Strands Agents SDK — an open-source agent framework. This is the core.
- Ollama — runs an open-source model (llama3.1) locally, so no API cost, no card, no bill.
Because Strands is model-agnostic, I could point the agent at a local Ollama model instead of a paid cloud one. That single design choice is what made this a ₹0 project.
Step 1 — The agent
Install Ollama, pull a model:
bash
ollama pull llama3.1Then the agent itself is tiny:
python
from strands import Agentfrom strands.models.ollama import OllamaModel
model = OllamaModel(host="http://localhost:11434", model_id="llama3.1")agent = Agent(model=model, system_prompt=SYSTEM_PROMPT)The whole "brain" is in the system prompt: explain in simple words, break down jargon, end with clear "what to do next" steps (including which documents to carry), and flag scams and deadlines. No wrong or scary advice.
One bug I hit: reusing a single cached agent threw "Agent is already processing a request" on quick clicks. Fix — build a fresh agent per request and cache only the model. Small thing, but it took me a minute to understand.
Step 2 — Make it usable by non-technical people
A plain text box isn't enough for elderly users. So I added three ways in:
- Type/paste — the basic path
- 📷 Photo — upload or take a photo of the message; EasyOCR pulls the text out
- 🎤 Speak — voice input via
streamlit-mic-recorder(speech-to-text), so no typing at all
And one way out that matters most: voice output with gTTS, so the answer is read aloud.
Step 3 — Their language, everywhere
If someone can already read English, they didn't need this app. So the whole interface — every label, button, the "what to do next" heading, and the answer — switches to the chosen language: Hindi, English, Marathi, Kannada, Tamil, Telugu, or Bengali. That was just a matter of passing the target language into the prompt and keeping a translation dictionary for the UI strings.
The stack, in one line
AWS Strands Agents SDK + Ollama (llama3.1) + EasyOCR + streamlit-mic-recorder + gTTS + Streamlit. All open-source, all local, ₹0.
What I actually learned
- You can build a genuine AI agent without a paid cloud model. Local models are further along than I thought.
- The Strands SDK's model-agnostic design is the real unlock — it let me focus on the problem instead of the plumbing.
- Starting from "I don't know AWS at all" is fine. The gap between that and a working agent was one focused day.
If you're staring at an empty AWS account feeling like everyone else knows more than you — start with a problem you actually care about, and build the smallest thing that works. That's the whole trick.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article