MSME-360: Building an AI Assistant That Remembers How Small Businesses Work

Reva Kale — Sat, 07 Nov 2026 13:53:42 +0000

By Reva Kale

I started this project thinking one retrieval system would be enough. I was wrong, and the mistake taught me what each tool is actually good at.

What We Are Building

MSME-360 is an assistant and dashboard for small businesses. It reads your messages, scans your receipts, creates GST invoices, and keeps your records in sync with Google Sheets.

I built the AI side: the chat, the search, and the memory.

The memory comes from Hindsight, an open-source agent memory system from Vectorize.

A few useful resources:

You can also explore our project:

Live Demo: https://pbl-project-tan.vercel.app/
GitHub Repository: https://github.com/reva-32/pbl_project
Project Demo Video: https://www.youtube.com/watch?v=x-LB7Hc1c6I

Two Kinds of Knowing

A business owner asks two very different kinds of questions.

The first is about facts in their records: "What was our highest invoice last month?"

The answer sits in a receipt or an invoice somewhere.

The second is about them: "Make this invoice like the last one."

The answer is not in any document. It is in the habits, decisions, and instructions the owner has given over time.

I tried to use one system for both, and the results were mushy.

Vector search is great at finding documents but has no idea what the owner said last Tuesday. Memory is great at remembering the owner but should not be the store for every receipt.

So I split the job.

Part One: Private Retrieval

Every uploaded receipt, bill, or invoice is turned into text, summarised, and converted into a 384-dimensional embedding.

These live on the DecisionEvent collection in MongoDB Atlas with a vector index called vector_index using cosine similarity.

When the owner asks a question, I compute an embedding for it and run $vectorSearch, always filtered by userId.

If the vector search comes back thin, a keyword regex fallback catches the obvious matches.

The top three records go forward.

Part Two: Hindsight Memory

Before the model answers, I recall relevant memories from Hindsight: past decisions, patterns, and instructions.

After the conversation, anything worth keeping is retained.

Over time, the assistant can also reflect on what it has stored to form a more useful picture of how the business runs.

Putting It Together

Here's how the complete flow works:

Question

↓

Embedding + $vectorSearch (userId filter) + keyword fallback

↓

Top 3 private records

↓

Hindsight recall + user profile defaults

↓

Groq model

↓

Answer, or an invoice PDF

The model gets documents that say what happened and memories that say how this owner likes things done.

What It Looks Like in Use

On day one, the owner asks for an invoice for Acme Corp for ₹25,000.

The assistant has to ask about tax and payment terms. The owner says to always add 18 percent GST and use 15-day terms.

That instruction is retained.

Days later, a different customer, a different amount. The assistant does not ask. It builds the invoice with the right tax and terms.

Without memory, every session starts from zero.

With memory, the assistant feels like it has worked here for months.

Models and Function Calling

We run on Groq, using models such as llama-3.3-70b and openai/gpt-oss-120b.

The invoice flow relies on function calling, and small models sometimes return a malformed call.

I added handling so that a bad tool call is caught and retried instead of showing the user an error.

If you build with tools, plan for this early.

What Was Painful

Deciding what to retain took longer than I expected.

If you retain everything, recall fills up with noise.

I now keep instructions, preferences, and business decisions, and leave raw document text to the vector index.

Lessons I Would Reuse

Separate "what is in my files" from "how do I work." Search and memory solve different problems.
Filter by user at the retrieval layer. Never trust the prompt to do isolation.
Keep a keyword fallback. Vector search alone misses exact names and numbers.
Be picky about what you retain. Quality of recall depends on quality of retention.
Expect tool calling to fail sometimes. Catch it and retry quietly.

The Tools

We ran Hindsight on Hindsight Cloud, adding the promo code MEMHACK99 in the billing section after registering for $50 of free credits.

We used the Band AI desktop app while building, and the Hindsight community Slack was a big help when we had questions.

Final Thoughts

Retrieval tells your assistant what is true.

Memory tells it who it is working for.

You need both.

DEV Community: Reva Kale