Skip to content

vishalsingh2972/telugu-thodu

Repository files navigation

Telugu Thodu (తెలుగు తోడు)

AI Voice Companion Infrastructure for Telugu Families

Telugu Thodu is an AI voice companion that calls Telugu parents living alone, understands how they are doing through natural conversations, and helps NRIs stay emotionally connected with their families back home.


Architecture Diagram

flowchart TD

A["📞 Parent Receives Scheduled Call"]

subgraph AI_Processing
    B["🎙️ Sarvam Saaras V3<br/>Speech-to-Text"]
    C["🧠 Gemini 2.5 Flash<br/>Conversation Analysis"]
    D["📋 Wellness Summary"]
end

subgraph Parent
    E["👵 Parent Speaks Naturally"]
end

subgraph Monitoring
    F["⚡ Upstash Redis"]
    G["📊 Wellness Dashboard"]
end

subgraph Alerts
    H["🚨 Urgency Detection"]
    I["📱 WhatsApp Alert to Child"]
end

subgraph Insights
    J["😊 Mood Score"]
    K["💊 Medication Tracking"]
    L["📈 Weekly Trends"]
end

M["❤️ Peace of Mind for Telugu Families"]

A --> E
E --> B
B --> C
C --> D

D --> F
F --> G

G --> J
G --> K
G --> L

D --> H
H --> I

L --> M
I --> M
Loading

🎥 Demo Video

Let's see Telugu Thodu in Action

2026-06-02.13-04-12_compressed.mp4

📌 Project Overview

Telugu Thodu (meaning Telugu Companion) is a proactive, voice-first wellness infrastructure designed for Telugu families separated by distance.

Instead of relying on manual check-ins or smartphone apps that elderly parents may find difficult to use, Telugu Thodu initiates automated voice calls directly to traditional Indian telecom lines and speaks naturally in Telugu/Tenglish.

The system understands conversational speech, extracts wellness signals like emotional stress or medication adherence, and delivers structured updates to children living abroad through a real-time dashboard.


👩‍💼 Real-World Example: Explaining Telugu Thodu to Telugu NRI Sudha living in Texas, USA

“Sudha, you're in Texas working long hours, dealing with time zones, and always worrying if your mom back in Hyderabad is genuinely okay. When you call her, she does what all Indian moms do — she hides her pain, says ‘Antha baane undi ra’ (Everything is fine), and hangs up because she doesn't want to stress you out. Telugu Thodu is an automated, friendly AI companion that calls your mom directly on her regular phone line at a scheduled time. It speaks to her in a warm, comforting, local Telugu accent. It asks how her day is going, if she took her blood pressure medicines, or if she has any pain. Your mom doesn't need a smartphone, an internet connection, or a tech degree. She just answers the phone and chats naturally, mixing Telugu and English like she normally does. The moment she hangs up, our AI listens to what she said, catches if she is secretly feeling weak or dizzy, and updates a clean dashboard for you. Instead of worrying or guessing, you look at your phone in Texas and instantly see a message: ‘Amma took her medicine but mentioned slight morning dizziness. She sounds a bit anxious today — consider giving her a call tonight.’ It gives you absolute peace of mind while keeping her safe without any tech hassle.”


💭 The Problem

For many Telugu NRIs living in the US or Europe, one silent anxiety never goes away:

“Are my parents actually okay?”

Parents often say:

  • “Everything is fine.”
  • “Don’t worry about us.”
  • “We’re managing.”

Even when:

  • medicines are skipped,
  • stress is increasing,
  • health issues are developing,
  • or loneliness is becoming severe.

At the same time:

  • time zones,
  • work schedules,
  • and physical distance make daily monitoring difficult.

Telugu Thodu acts as an empathetic AI companion that bridges this emotional and informational gap.


👵 Parent Experience — How Telugu Thodu Helps

The parent receives a simple phone call.

No app installation.

No typing.

No smartphone knowledge required.

Just a natural conversation like:

“Namaste amma, medicines teesukunnara today?” “Ela unnaru?” “Tindi time ki tinnara?” “BP or sugar issue emaina unda?”

The parent simply speaks normally in Telugu or Tenglish.


🧠 Cognitive Processing Pipeline

The AI system analyzes conversations for:

  • emotional tone,
  • stress indicators,
  • loneliness signals,
  • health complaints,
  • medicine adherence,
  • conversational irregularities,
  • urgency detection.

Example

If a parent says:

“Konchem dizziness undi… medicines miss ayyayi.”

The system may infer:

  • possible health discomfort,
  • skipped medication,
  • follow-up recommendation.

📱 NRI Dashboard Experience

Instead of forcing children abroad to listen to lengthy call recordings, Telugu Thodu compresses raw conversational data into structured wellness summaries.

Analytical Matrix Extracted Value
Mood Index ⚠️ Slightly Anxious
Medication Adherence ❌ Skipped
Reported Issues ["Dizziness", "Weakness"]
Action Required TRUE
AI Summary "Mother sounded weak and mentioned dizziness after skipping blood pressure medication."

Within seconds, NRIs can understand how their parents are doing emotionally and physically.


🛠️ Tech Stack & Engineering Rationale

Architecture Layer Technology Engineering Selection Reason
Framework & Router Next.js 15 (App Router) Full-stack TypeScript architecture enabling asynchronous serverless webhook execution and modern streaming UI rendering.
Data Integrity Layer Zod Validation Enforces strict runtime parsing and schema safety across unpredictable AI-generated outputs.
Telephony Middleware Twilio Programmable Voice Handles telecom edge delivery, TwiML execution maps, recording buffers, and asynchronous call orchestration.
Sovereign Speech AI Sarvam AI (Saaras V3) Configured explicitly for Speech-to-Text (mode: "codemix", language_code: "te-IN") to transcribe and translate raw Telugu/Tenglish conversational audio into clean medical English.
Inference Engine Gemini 2.5 Flash Used for rapid structured semantic analysis and deterministic object generation.
State Cache Layer Upstash Redis Maintains transient telephony lifecycle states (CALL_TRIGGERED, IN_PROGRESS, ANALYZING) with ultra-low latency.
UI Presentation Layer shadcn/ui + Tailwind CSS Utility-first responsive interface for clean dashboard rendering.

🔄 End-to-End Telephony Lifecycle

The system decouples unstable telecom behavior from internal processing logic using an asynchronous event-driven architecture.

[Parent Mobile Device / Telecom Carrier]
                    │
                    ▼ (8kHz Telephony Audio)
     [Twilio Voice Infrastructure]
                    │
                    ▼ (Encrypted Webhook Tunnel)
     [Next.js Serverless Processing Layer]
                    │
                    ├─► [Sarvam Saaras V3 API]
                    │         └─► Telugu/Tenglish Speech-to-Text & Translation Translation
                    │
                    ├─► [Gemini 2.5 Flash]
                    │         └─► Semantic Analysis + Structured Evaluation
                    │
                    ▼
        [Zod Runtime Validation Layer]
                    │
                    ▼
        [Upstash Redis State Cache]
                    │
                    ▼
    [Real-Time Wellness Dashboard]
                    │
                    ▼
     [Weekly Scheduler / Digest Engine]
                    │
                    ▼
[WhatsApp API (NRI Child Abroad)]
                    └─► Weekly Wellness Summary Report

📋 Telephony State Machine

Lifecycle State Trigger Vector System Action Output
CALL_TRIGGERED Cron / Manual Trigger Initialize Twilio session & allocate Redis tracking key callSid, status
IVR_STREAMING Parent Answers Dispatch Telugu greeting & initialize recording flow TwiML XML
STT_PROCESSING Call Ends Stream .wav buffer to Sarvam Saaras V3 Raw Tenglish Transcript
COGNITIVE_ANALYSIS Transcript Ready Execute semantic wellness analysis Structured JSON
STATE_RESOLVED Validation Success Persist validated object & refresh dashboard WellnessReport
WEEKLY_DIGEST_DISPATCHED Weekly Scheduler Trigger Generate longitudinal wellness summary & deliver update to NRI child via WhatsApp WhatsApp Wellness Digest

The system never trusts raw LLM output directly.

Every AI-generated response is forced through deterministic runtime validation before touching application state or UI rendering.

import { z } from 'zod';

export const WellnessReportSchema = z.object({
  overallMood: z.enum([
    'positive',
    'neutral',
    'anxious',
    'depressed',
  ]),
  reportedHealthIssues: z.array(z.string()),
  medicationStatus: z.enum([
    'taken',
    'skipped',
    'unknown',
  ]),
  actionRequired: z.boolean(),
  summaryForChild: z.string().min(10).max(250),
});

export type WellnessReport =
  z.infer<typeof WellnessReportSchema>;

🚀 What I Learned from this Project:

  • Built my first AI-powered voice application to solve a real-world problem.
  • Learned how to integrate Twilio telephony, Sarvam AI, and Gemini into a single complete workflow.
  • Explored Telugu/Tenglish speech transcription and multilingual AI systems for the first time.
  • Gained experience designing event-driven and serverless system architectures.
  • Understood how multiple services can come together to solve a meaningful problem for families.
  • Got ands-on experience building and deploying a full-stack AI product from idea to working prototype.

🌏 Why This Project Exists

Telugu Thodu is not trying to replace family relationships.

It exists to strengthen them.

The goal is simple:

Use localized AI voice infrastructure to help Telugu families stay emotionally connected across continents.


About

Helping Telugu NRI remain fully connected to family in India ❤️

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors