Bookings and appointments
Patients into a clinic system like Simplex. Venues, trips, and visits, from the request to the confirmed record and the reminder.
How a process
stops needing people.
The same four parts, whatever the process is. Below: where people get stuck at each one, what I build there, which processes fit the pattern, and what I am still working out.
- Intake
- Where people get stuck
- Work arrives as a phone call, a WhatsApp message, an email, or a form, and someone has to read it and type it somewhere.
- What I build
- Agents on every channel the work comes in on, in English and Arabic, that understand the request and pull the context they need.
- Decide
- Where people get stuck
- The rules live in someone’s head. The system either guesses or asks a person every time. Performance is judged once a quarter from memory.
- What I build
- Explicit decision logic: what the model decides, what a rule decides, and the exact conditions under which a human is asked. Scores and decisions come from the work record, not from a form.
- Act
- Where people get stuck
- The decision is made, and now someone still has to put it into the calendar, the CRM, the clinic system.
- What I build
- Tools that write directly into the systems you already run, with identity checks and permissions inside the tool.
- Recover
- Where people get stuck
- Something fails at 2am and nobody finds out until a customer complains.
- What I build
- Failed actions queue and replay. State survives interruptions. The system reports what it could not do, so the exceptions list is the only thing a person reads.
What I automate.
Anything a business does the same way twice.
The pattern is always the same: find where work enters, where a decision is made, where a system gets updated, and where things break. Then remove the person from everything except the exceptions.
Sales and customer conversations
Every channel a customer uses, with the answer, the booking, and the CRM update handled in the same flow.
Hiring
I built the Recourse Loop applicant tracking system with Outlook scheduling by hand at Xloop. The automated version has the same shape: screening, scheduling, follow-ups, with a person only for the final decision.
Employee performance
Scored continuously from tasks, submission times, QA results, and live feedback, the way Track Loop does it. No review forms, no quarterly guesswork.
Operations inside the team
Task allocation from incoming work, approvals, and the follow-ups that keep a project moving without a manager chasing.
Events and access
Invitations, RSVPs, QR admission, and staff coordination, with state that survives changes mid-event.
Whatever you do by hand
If it happens the same way every time and a person still has to be there, it fits the pattern. Tell me the process and I will map it.
What I measure on every one: the handover rate. How often a person had to step in, and why. That number is the product.
What I’m still
working out.
The open questions in automating
whole processes, and where I am with them.
When should the system ask a human?
Too early and you have automated nothing. Too late and you have a mess. I design the handover conditions explicitly for each process, and I keep looking for a general rule.
How do you test a process, not a function?
Unit tests do not catch a model that starts deciding differently. I build evals around the business outcome, and I am still learning which ones predict trouble.
What does an exceptions list need to say?
If a person only reads what the system could not do, that list is the whole interface. What it says, and how fast a person can act on it, decides whether the automation holds.
Can a model judge a person’s work fairly?
Track Loop scores from tasks, timing, QA, and feedback, never from hours alone. The open question is which signals predict real performance and which just reward being fast. I keep the weights visible so a manager can argue with them.
The stack, by part.
What each part of an automated process is built with.
- Intake
- VAPI, Hamsa, Twilio, WhatsApp, Instagram, Messenger, SMS, Gmail, Outlook
- Decision
- Azure OpenAI, retrieval pipelines, tool schemas, explicit rules, evals
- Systems it writes into
- Simplex, Odoo, Bitrix24, Google and Microsoft calendars
- Backend and data
- Node.js, NestJS, RabbitMQ, WebSockets, PostgreSQL, MongoDB, SQL Server, AWS S3
- Interface
- TypeScript, React, Next.js, Tailwind CSS
A few things
you might ask.
What kind of role are you looking for?
A senior engineering role where the job is to take real business processes and make them run without people. I’m most useful where the system has to act in other systems, not just answer questions.
Does “automate everything” mean no humans at all?
No. It means a person handles exceptions the system flags, instead of every step. Deciding exactly when to ask a human is part of the design, and I treat the handover rate as the main measure of the work.
Can AI really score employee performance?
It can score the work. Track Loop reads what was delivered, when, how QA rated it, and what the client said, and keeps a running score. A manager still owns the conversation and the final call. The system removes the form-filling and the guesswork, not the manager.
Do you only do the AI part?
No. I build the interface, the backend, the queues, and the integrations around the model. The model is one part of the process.
What’s the best way to start?
Email me with one process your team does by hand, from where the work comes in to where it ends up. That gives us something concrete to talk about.
hunain.parekh@hotmail.com