I build AI systems that run in production.

Peak call volume overlaps with peak dining hours.

During dinner service every host is on the floor. The phone goes unanswered, callers don't leave voicemails, and reservations walk out the door. A dedicated phone host for every shift isn't economical, and traditional IVR frustrates callers.

A voice AI that knows the menu and hands off when it matters.

• Answers hours, menu, FAQ and reservation questions from a live knowledge base
• Three routing paths by party size and timing: callback, online booking, or the events form
• A nightly n8n job updates date awareness and republishes the agent
• Every completed call is turned into a WhatsApp and email alert within 2 minutes

Every call answered, every reservation captured.

Calls that used to fall through during service now reach Maya. Name, party size, date, time, email and occasion reach the team within 2 minutes, and confirmations or complex requests route straight to a human.

Manual ad review didn't scale, but full autopilot was too risky.

Weekly review across multiple restaurant ad accounts meant cross-referencing pixel data, ad status and budgets by hand: slow, and easy to defer. Letting an AI change live ad spend unsupervised was not an option.

Claude analyzes, a human approves, nothing silently auto-applies.

• Claude reviews pixel, ad and budget data every Monday and proposes specific changes
• Approve or reject by email link or the dashboard; the first response wins
• Granular selection: individual ads, ad sets or conversions
• 48-hour timeout defaults to reject
• Now on Google Ads too, with the same decision format and validate-only dry runs

AI does the analysis; the operator keeps the final call.

Runs in production every Monday across both accounts, with budget changes (±20%) applied only within approved selections. A separate daily monitor flags spend spikes (≥1.5x trailing average) between the weekly cycles.

Two sources of truth, checked by hand, if at all.

Staff record cash due and net sales per shift in the Ops Console. The POS emails its own close-out report the next morning. Matching them meant opening a PDF and adding up checkouts by hand.

Parse defensively, compare server-side, keep a human override.

• Blank cells vanish from the PDF's text, so the parser finds the split where each row's own arithmetic adds up
• The report's send date is shifted back to the business day it covers
• The comparison runs inside the Ops Console, which stays the source of truth
• Admins can mark a known difference as verified, with a note

Every day starts with a verified or flagged number.

Each day's report shows a green check or a red flag for cash and net sales. Red is the default until a check passes, so an unverified day never looks fine by accident.