Babu, an elderly-companion assistant
A Hindi-speaking companion for older adults that remembers their life and keeps them connected.
1 · Purpose
Many older adults spend long hours alone, and generic voice assistants forget everything between conversations. Babu is a companion that knows the person over time: their life story, the people they love, their routines and what is happening in their family's lives. At its core is a personal world model, a structured memory of people, events, facts, preferences and time, rather than a plain chat log. The work is requirements-first: the requirements, memory design and architecture came before any technology choice, and the first build is a prototype on a laptop before an Android tablet app.
2 · Business benefits
3 · Status and how to join
Good first tasks this week:
- Write a Hindi or Hinglish exchange for one everyday situation, such as a festival morning.
- Review a sample transcript and mark anything that sounds like a machine.
- Add a test case for who-does-he-mean, such as two people with the same name.
4 · Requirements
What it must do
- Hold a natural conversation in Hindi, English or a mix of both.
- Remember people, relationships, events, preferences and routines, and bring them up naturally later.
- Find memories four ways: by exact facts, by meaning, by relationships and by time, then combine the results.
- Forget something completely when the person asks, including every copy and index of it.
- Handle sensitive moments, such as a bereavement, without saying anything false.
- Work out who the person means and what they are asking, and ask back when it is unclear.
- Offer gentle activities such as songs, proverbs and riddles.
How well it must do it
- The older adult is the owner of their own data; family members cannot consent for them.
- Built to meet India's data protection law, including deletion that really deletes.
- Private conversations are never sent to a model provider that may train on them.
- Replies come quickly enough to feel like a person thinking, not a machine loading.
- Large text, clear voice and no timers or menus the person must race against.
- Every remembered fact keeps who said it and the original words in the original language.
5 · Architecture, very high level
6 · Tech stack, per platform
| Component | Open stack | Azure | AWS | Android |
|---|---|---|---|---|
| World model store | PostgreSQL with pgvector, relationships as rows and recursive queries ✓ | — | — | — |
| Meaning search | Multilingual embedding model (BGE-M3) ✓ | — | — | — |
| Schema changes | Alembic migrations ✓ | — | — | — |
| Tests | pytest with guarantee, retrieval and tone suites ✓ | — | — | — |
| Reply model | Proposal: a hosted model with no-training terms, or self-hosted open weights | — | — | — |
| Voice | Proposal: speech-to-text and text-to-speech around the same chat loop | Proposal: Azure AI Speech for Hindi | Proposal: Amazon Transcribe and Polly | — |
| Store | — | Proposal: Azure Database for PostgreSQL with pgvector | Proposal: Amazon RDS for PostgreSQL with pgvector | — |
| Model | — | Proposal: Azure-hosted model with no-training terms | Proposal: Amazon Bedrock | — |
| App | — | — | — | Proposal: a tablet app with large text and voice first, after the laptop prototype |
| On-device voice | — | — | — | Proposal: Android speech services, only if they pass a Hindi quality test |
7 · What needs to be done
Requirements and design 3 of 3 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Write and review the requirements catalogue, with a test for every must-have | done | L | requirements |
| Design the memory ontology and canonical schemas | done | L | data modelling |
| Choose the store and write the architecture and database design | done | M | architecture |
Memory core prototype 4 of 4 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Schema, write path and true deletion, with tests | done | M | PostgreSQL, Python |
| Seed a realistic fictional life of about 200 records | done | M | writing, data |
| Four-way retrieval, permission gate and conflict handling | done | L | search, Python |
| Understanding step: who is meant and what is asked | done | M | NLP |
Conversation prototype 0 of 3 done
| Task | Status | Size | Skill |
|---|---|---|---|
| One-hour tone probe on real model replies | todo | S | conversation design |
| Prompt assembly, reply generation and memory keeping | todo | M | LLM engineering |
| Tone harness that fails the build on banned phrasing | todo | S | testing |
Voice 0 of 1 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Add speech in and out around the chat loop and test Hindi voice quality | todo | L | speech |
Android tablet MVP 0 of 1 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Build the tablet app on the proven prototype | todo | L | Android |
8 · Roles
| Role | How many | Hours / week | What they do |
|---|---|---|---|
| Conversation designer | 1 | 2–3 hours/week | Writes and judges Hindi exchanges and the tone rules. |
| Backend engineer | 1 | 4–6 hours/week | Builds the reply pipeline and memory keeping. |
| Speech engineer | 1 | 2–4 hours/week | Tests and integrates Hindi speech in and out. |
| Android developer | 1 | 4 hours/week | Builds the tablet app once the prototype is proven. |
9 · What you'll learn
- How to design long-term memory for an AI assistant beyond a vector database.
- How to combine search by facts, meaning, relationships and time.
- How privacy law shapes a data model from the first record.
- How to design conversation for older adults and sensitive moments.
- How to work requirements-first and choose technology last.
Level-up questions
- Does Babu sound like a person rather than a form when you read its transcripts?
- Is on-device Hindi speech good enough to live with, or must voice run in the cloud?
- How should Babu handle a family member asking about something the person chose not to share?
- What is the smallest set of situations the prototype must handle beautifully?