An owner working between interruptions.
Customer questions arrive while the owner is already handling products, prices and documents. The useful tool has to fit that working day.
Case study · small-business agent
Kelly
A voice counter assistant, built for two real shops: an electrical shop and my mother's boutique, She Fashion House. Customers talk to Kelly hands-free. She understands Hindi, Hinglish and English, and always answers in English, showing designs and putting together quotations with prices computed in code.
Codex-only configurablelocaldemo-readyhands-free, English replies
Live test build
Talk to Kelly
My voice assistant for shop counters. She runs on my Mac, so she is only awake when it is.
Talk to Kelly Checking if she is awake...
Kelly runs on my Mac. Try again later.
Ask me for a guest login.
the person and the problem.
Kelly started with my cousin's shop, and now runs the counter for an electrical shop and my mother's boutique, She Fashion House. The work was familiar: find the right item in a catalogue, confirm its price, apply GST, prepare a quotation and answer the next customer question without losing the thread.
The answer lives across catalogues and calculations.
A quotation is not one prompt. It needs the right product record, price, GST treatment, a usable document and a final human check.
the product decision.
I did not build Kelly as a generic chatbot or as software only for electrical shops. I built a configurable agent around the business's own catalogue and the jobs people already do: quotation, catalogue lookup and customer support.
The decision: keep the agent local, make Codex the only configurable reasoning layer, let people speak hands-free in Hindi, Hinglish or English, always answer in English, compute prices in code rather than in the model, and stop for owner review before action.
how it works.
import the catalogue
Bring product data in from PDF, XLSX or CSV and publish it for retrieval.
ask naturally
Talk hands-free in Hindi, Hinglish or English, or use web, terminal or Telegram. Kelly always answers in English.
build the answer
Retrieve the product, show matching designs, compute pricing and GST in code, then prepare the quotation or support response.
review before action
Show the result to the owner. The human decides what is ready to use.
what I owned.
- Product framingStarted from my cousin's shop workflow, then designed Kelly for quotation, catalogue and customer-support jobs.
- Catalogue pipelineImported PDF, XLSX and CSV data, then grounded retrieval in the published catalogue.
- Quotation workflowConnected product retrieval with pricing, GST, document preparation and spreadsheet work.
- Conversation designAdded hands-free Hindi, Hinglish and English voice input with replies always in English, plus isolated Q&A conversation RAG.
- Working surfacesMade the core workflow available through terminal, web and Telegram.
- Human controlKept owner review as the decision point before action.
current stage and next improvement.
Demo-ready.
Kelly runs locally today with catalogue imports and retrieval, pricing and GST, quotations, spreadsheet workflows, voice and text surfaces, and owner review.
She is reached through a Cloudflare tunnel from my Mac, so she is online only when the Mac is.
Make setup easier for each business.
The next step is a clearer configuration flow for catalogue fields, GST rules, quotation preferences and support behaviour, so a new business can adapt Kelly without changing its core workflow.