Bala Pothabattula

What makes it work

• Nothing sends automatically — the system has no permission to send mail at all. Every reply is left as a draft for the lawyer to review, edit and send. That's a hard guardrail for a regulated practice, not a setting.
• Two AI models split the work by cost — a fast, cheap one triages every email; a more capable one writes the drafts that actually need care.
• It drafts in the lawyer's own voice, learned from samples of their real writing — so a draft reads like them, not like a chatbot.
• Every firm's data is isolated by row-level security and the mailbox token is encrypted at rest — built multi-tenant so more firms can be added without ever mixing data.
• Every email it touches is logged with a full audit trail and exportable to CSV — so the firm can show exactly what the assistant did and when.

What it doesn't do yet

• A feedback loop — thumbs up/down on each decision so it learns the lawyer's preferences over time — is specced and partly built, not yet live.
• Attachment reading covers normal PDFs and Word files; scanned documents are flagged but not yet read (OCR is a later phase).
• VIP-sender alerts and a first-run pass over older email are designed but not switched on yet.

What makes it work

• It only quotes hotels and prices it actually retrieved — it can never invent a property or a rate, and it never reveals the company's margin, even when a customer asks directly.
• A strict 50-rule guardrail keeps it on task: it promises a price only after confirming dates and room, resists prompt-injection attempts, and escalates complaints, refunds and anything high-stakes to a human.
• Discounts are decided by the server, not the AI — the agent can only offer a reduction within a margin floor the business sets, so it can never give away more than the business allows.
• Supplier and payment connections sit behind swappable adapters — the client's live Rezlive and Nomod accounts drop in without touching the conversation engine.
• A live admin dashboard shows every conversation, flags escalations and abandoned payments, and lets a colleague join and take over any chat from the agent.

What it doesn't do yet

• Live supplier and payment connections — it runs on demo hotel inventory and simulated payment links today; swapping in the client's real Rezlive and Nomod accounts is the next step.
• WhatsApp delivery is wired to the Meta Cloud API but not yet tested on a live number — that switches on once the client's WhatsApp Business account is verified.
• Conversations are held in memory for the demo; moving them to a shared database is needed before running across more than one server instance.
• Multi-item carts and a gentle nudge for customers who start but don't finish a booking are planned, not built.

What makes it work

• Two AI models work together — a fast one runs the live conversation in real time, a more thoughtful one writes a plain-English summary of the call for the owner afterwards.
• It recognises regular customers by phone number and greets them by name.
• It checks the calendar live before quoting any slot — so it never books a time that conflicts with an existing appointment.
• Allergy questions, complaints, large-ticket enquiries, and anything the AI isn't confident about transfer straight to the owner — it doesn't pretend to handle things that need a human.
• Every tenant runs in complete isolation — separate RLS context, separate conversation history, separate per-tenant cost caps. One business's calls never touch another.

What it doesn't do yet

• Sharper flagging of urgent items at the top of the morning call summary — currently all items carry equal weight.
• Broader calendar provider support — Google Calendar and Outlook covered; iCal sync is on the roadmap.
• Outbound appointment reminders — the infrastructure is in place; it's a configuration step per tenant.

What makes it work

• The business's own QR code is sent with every order — we attach the image only. No payment processing, no holding of funds, no PCI scope.
• Allergy questions, complaints, refunds, and any high-stakes situation trigger an instant transfer message to the owner — a regex pre-filter catches these before the LLM even reads the message.
• Each conversation is isolated by tenant via Postgres RLS — one tenant's customer history is never visible to another tenant's agent context.
• Per-conversation and per-tenant cost circuit breakers cap spend automatically — if usage spikes unexpectedly, it pauses and alerts rather than running up an unbounded bill.
• Returning customers are recognised by phone number and greeted by name — no extra setup required.

What it doesn't do yet

• Richer multi-item order carts — the current flow handles a single order item per conversation cleanly; grouped multi-item orders are on the roadmap.
• Automated follow-up nudge for incomplete orders — customers who start a conversation but don't confirm; a gentle 30-minute reminder is planned.
• More verticals — the system is vertical-agnostic but prompt configuration needs calibrating per business type; currently optimised for the six target verticals above.

What makes it work

• Pulls leads from official UK sources — Companies House, Google Maps, healthcare and veterinary registers — so the businesses are real, not scraped contact data.
• A safety check runs on every contact before anyone gets approached — no chasing people you've already spoken to, no contacting blocked names, no wrong-time outreach.
• The call brief is written by the same AI tools used to power the most reliable enterprise software today. It reads the website and the reviews and writes a tailored opener.
• I deliberately built only the research half. I didn't build the auto-emailing half — most businesses already have email tools they like, and choosing the wrong one for them would cost more to undo than to skip.

What it doesn't do yet

• Email and CRM integration are not included by design — those steps tie tightly to your existing tools, and the wrong fit early would be expensive to undo. They're a natural next step once you've picked your stack.
• The 'type of customer' definition is currently a sensible UK default — first thing we'd customise for any new client.

What makes it work

• It only reads public sources — no scraping, no privacy issues, no legal risk.
• A pre-filter throws out 60–70% of the noise before the AI even reads the post — so the system doesn't waste money on irrelevant content.
• It groups similar complaints together, so you don't see the same problem ten times in a row.
• If a topic suddenly spikes — five people complaining about the same thing in a single day — it sends an alert.
• The end-of-day brief is an actual one-pager in your inbox or Slack — not another dashboard you have to remember to check.

What it doesn't do yet

• The dashboard view (where you can browse by topic) currently shows sample data while we focus on getting the daily brief right. The plumbing is built; we'll switch it on once the daily brief has earned its keep.
• Two extra features are designed but not yet running: a deeper competitive analysis when a topic gets interesting, and an auto-drafted social-media post about findings. Both are deliberately on hold until we're confident the daily brief is the right shape.

What makes it work

• Findings are ranked by what they're costing your business in pounds, not by technical severity. The report has to make sense to your finance director as well as your engineer.
• I never modify anything during the audit — it's read-only. If you want me to fix what I find, that's a separate engagement with explicit sign-off.
• Every finding has a 'if we don't fix this, what happens' line. Severity without consequence is just opinion.
• The cost review uses AWS's own attribution data, not just the bill — so we can see exactly where money is leaking.

What it doesn't do yet

• Currently it's all manual delivery. I'd like to automate the easy diagnostics so you get the boring findings in the first day, leaving the rest of the week for the parts that need real judgment.
• I'm considering open-sourcing the diagnostic checklist so any business can run a basic health-check themselves.