Longbook

Inspiration

Small business owners often juggle scattered records - spreadsheets, WhatsApp chats, invoices, even voice notes but don’t have the time or expertise to turn them into meaningful insights. The goal was to give them the same analytical power that large enterprises enjoy, but in a way that feels simple and accessible.

The idea also comes from a personal pain point: many small business leaders are “one‑man armies,” managing sales, operations, and finance all at once. It’s overwhelming, and we wanted to build a digital right‑hand assistant that lightens that load.

At its core, the app is about democratizing data. Data‑driven decision‑making shouldn’t be reserved for big corporations; it should be intuitive and self‑configuring for every trade, whether you’re a textile trader or a local retailer.

Finally, trust was a key motivation. Unlike black‑box dashboards, every insight in our app is traceable back to its source record. That transparency builds confidence, ensuring business owners know exactly where each figure comes from.

What it does

Longbook is a self-configuring operations layer specially created for small businesses to help them track what’s happening with their business and provide them necessary insights for the next apt move. It reads the records a business already keeps like email, WhatsApp exports, excel and Tally sheets, photographs relevant to business context, spoken notes and helps to provide information on potential customers, orders, counterparty tracking and analytical deep dive into various business aspects available to it.

How we built it

FastAPI, Postgres and LangGraph on Google Cloud Run, with a Next.js dashboard, all running in the Mumbai region. Ten agents. Six call Gemini; four are deliberately deterministic, because party resolution, the auto-commit threshold, ageing arithmetic and duplicate detection must be repeatable rather than probabilistic.

The pipeline groups messages into conversation windows, extracts a candidate record from each window, resolves it against existing parties, then either commits it or sends it to a review queue. Every window carries a content hash, so re-reading a history costs a model call only where something actually changed.

Challenges we ran into

Extraction failed structurally at first. Reading one message at a time, ninety percent of records could not be committed, because a trade order is built across several messages and a single line is not a record. Changing the unit of extraction from the message to the conversation window took auto-commit from ten percent to seventy-eight, without lowering the confidence threshold.

Gating a whole record on one uncertain field also forced owners to re-enter everything. Moving to field-level confidence meant the record commits what it is sure of and asks only about the field it is not.

Several failures only appeared in production. A backfill running inside the API process died silently whenever a deploy landed mid-run. Media sent to Gemini as a storage URI was rejected outright, and the failure lived in a process the dashboard could not see, so uploads looked successful while extracting nothing.

Accomplishments that we're proud of

The configuration agent generates a different interview for every business from that business's own records. A spare parts supplier is asked whether job work is billed separately, quoting his own messages about knurling and bobbins. A garment seller is asked what he calls his product codes, quoting the codes themselves. Neither set of questions could be asked of the other business, and neither is expressible as a template.

We also held a line under deadline pressure: the confidence floor was never lowered to make a metric look better. The chat section performed better than expectations handling huge business memory data with high accuracy.

What we learned

Measure before tuning. An early prompt change raised field accuracy while type accuracy stood still, and the averages hid regressions. Without an evaluation harness we would have shipped it as an improvement.

And nothing substitutes for a real customer. Our first pitch produced the most useful feedback we have received: exporting chats one contact at a time is impractical, which is now the next thing we are building.

What's next for Longbook

We are looking to increase our online presence via different marketing channels and technically, continuous ingestion. Email already syncs automatically; WhatsApp is still export-only, and live sync through Meta's Coexistence feature is the next build. Beyond that, multi-user access for the manager and the accountant, and reconciliation back into Tally.

Built With

• claude
• fastapi
• gemini
• google-cloud
• langraph
• python