A voice and chat assistant for Miraella, my merch business—retrieving customer history and creating drafts I can review before confirming.
My contribution
I defined the workflows, designed Covey’s conversational behaviour, configured the no-code prototype, and tested it. Prompt drafting and setup were AI-assisted. Portfolio development continued after the initial v1.
01 / Problem
Miraella sells personalised photocards, polaroids, posters, and stickers. Customer information lived across WhatsApp, Instagram, Google Forms, and handwritten notes.
A returning customer / Meera
When Meera returned to order customised polaroids through WhatsApp, I scrolled through our conversation for previous review photos and asked which freebie she had received before. To personalise her next order, I was reconstructing information from a relationship I already had.
Reconstruct the customer’s context for each repeat order
02 / Solution
Covey turns a conversation into a draft in Airtable. I can review and correct the details before confirmation. Capturing information and accepting it as history are separate steps.
Working demo
Recording to add
03 / Core flows
01 / Capture an order
Draft an order from natural language, linking it to the right customer. For someone new, create a customer-profile draft as part of the workflow.
Transcript excerpt / spoken request
“Eight… actually, make that six.”
A correction within the same utterance.
Reported draft result
The spoken correction was retained in the draft.
02
Bring past orders and confirmed packing updates into the conversation, so a repeat order starts with context already recorded.
Answer and source-order excerpt pending · Match the date and title.
03
Record what actually went into the package against the correct order. Review the packing update before it becomes confirmed history.
Record excerpts pending · Show requested and actually packed freebies separately.
04 / Design decisions
A gift recipient’s interests belong to that order; they should not automatically become the buyer’s preferences.
A request records intent; a packing update records what happened. Keep freebies separate from purchased items and preserve both stages.
Drafts stay out of confirmed-history searches until I review and confirm them.
05 / Testing & iteration
Two before/after comparisons reconstructed from test summaries, plus a wording issue still awaiting its empty-result retest. Original records and full logs are pending.
My decisionRevise extraction instructions to distinguish purchased items from freebies.
Before
After
Add the original draft, revised instruction excerpt, and matching retest capture.
My decisionFix the response mapping; retain honest uncertainty when success cannot be confirmed.
Before
After
Add the uncertain response, mapping change, and successful creation log.
Response wording
“nothing recorded yet”
Revised meaning: No confirmed history was found. That does not establish whether a draft exists.
Exact revised response pending · Empty-result retest pending.
Add the original response and revised wording, then run the empty-result retest. Keep the successful confirmed-history retrieval as separate evidence.
06 / Working system
ElevenAgents handles the conversation, Make connects the tool workflows, and Airtable holds the records. Customer profiles link to orders; packing updates link back to their order.
Conversation & tool calls
Workflow orchestration
Linked customer records
Airtable / record relationships
07 / Results & reflection
These behaviors were observed in prototype tests. They are examples of working paths, not a measured success rate or production performance claim. The empty-result wording retest remains pending.
Observed in prototype tests
Supporting scope
Compare time to the correct answer on the same tasks. Report the median and task count.
Check quantities, recipient, items, and freebies before editing. Count unsupported additions as errors.
Count correct links, drafts, and retrievals across every attempted scenario, including failures.
The key design lesson: useful customer memory depends on distinctions that conversation can blur—who bought, what was requested, and what actually happened. Making those distinctions explicit is as important as making capture easy.