Angelo Torrevillas · AI Automation Specialist
I build AI and automation systems that reduce repetitive work, such as following up overdue invoices, answering routine customer questions, preparing proposals and keeping your CRM up to date. I connect the tools your team already uses, use AI only where it helps, and build processes that are more dependable.
Services
Most of the work I automate falls into four areas. The examples below are labelled honestly. Most are portfolio demonstrations built with sample data or test payments, and one was delivered in a paid client trial. None is presented as a live client deployment.
Chasing unpaid invoices, following up leads and reminding people about bookings is repetitive and easy to get wrong.
I build automations that send the right message at the right time, stop once someone has paid, booked or replied, and set unusual replies aside for a person to review.
Related projects
Staff and customers ask the same questions again and again, and the answers already exist in your handbooks, service documents or records.
I build assistants for Slack, Facebook Messenger or the web that answer from the documents and data you provide.
Related projects
Copying invoice details into a spreadsheet, pulling requirements out of tender documents or preparing proposals by hand is slow, and small mistakes are costly.
I build systems that read the document or form, pull out the details, apply your rules, and either save a checked record or prepare a draft for a person to review and approve.
Related projects
When enquiries, bookings and payments live in separate tools, leads go cold and staff update records by hand.
I build workflows that move contacts through your sales pipeline, send booking reminders and confirmations, check online payments before access is released, and alert staff when something needs them. When a website or landing page is part of that journey, I build it as part of the workflow.
Related projects
Not sure which of these fits? Book a call and we'll look at your process together.
Featured projects
Three larger systems, built as portfolio projects using sample data or a fictional business. Each case study explains the business problem, how the system handles it, where a person stays in control, and what has and hasn't been tested yet.
Follows up overdue invoices by email, with one message per customer, firmer as the invoice ages, and a final payment check just before each email is sent.
Helps a bid team answer tenders by pulling out the requirements and drafting answers from previously approved content, with a person approving each answer and signing off the bid.
A fictional clinic's website and patient journey, designed so patients choose a doctor, pay online and receive their consultation link only after the payment is confirmed.
Client work
Delivered during a paid trial for an interior design firm. The firm isn't named, and its pricing rules, proposal content and client details are withheld.
Client details withheldTurns a sales enquiry into a priced proposal PDF. Fixed rules calculate the base fee, AI reviews the proposal and recommends the final fee, and Sales reviews the figures before the proposal is built and approved for sending.
All builds
Smaller builds and experiments, including chat assistants that answer from documents, a voice booking assistant, invoice data capture and CRM follow-up pipelines built in n8n, Zapier and GoHighLevel. Several ideas in the larger systems above were first tried here.
Answers staff questions about company policy in Slack, using the company's own handbook documents. Built and tested with a sample handbook.
Replies to customer questions on a business's Facebook page, using a document that describes its services. Built and tested with test data.
Send a photo of an invoice to a Telegram chat. It reads the details, catches missing key details and duplicates, and saves a checked record. Built and tested with test invoices.
Ask questions about sales data in plain English and get answers with the figures. Tested on a sample dataset. AI answers can be wrong, so check important numbers.
An AI phone assistant that lets callers check, book, move or cancel appointments in a clinic's calendar. Tested with test calls only.
Tests the simplest way to give a chat assistant reference material, by writing it straight into its instructions.
Tests loading two full reference documents on every message, which gives complete context at a higher cost per message.
Tests letting the assistant decide for itself when to open the reference documents.
The first version of the sales assistant, answering from a small sample sales table. It was later replaced by the SQL Sales Analytics Assistant.
When the team moves a lead to a new stage in Asana, the matching follow-up runs automatically, from project folders and tasks to timed emails.
Looks up company details for each new form enquiry, scores the lead with AI and sends a different follow-up depending on the score.
Turns one audio recording into blog and social media drafts, then posts them to Facebook and LinkedIn or holds them for a person to review. Connected to my own social accounts.
Moves a car insurance lead from first contact to an issued policy, with the right email, task and team alert at each stage.
Takes a salon client from first enquiry through their appointment to a rebooking offer, with emails, notes and staff tasks at each stage.
Follows up new leads until they book an appointment, then confirms the booking and welcomes paying clients.
15 builds
Process maps
Each map shows a process from start to finish in plain business terms, including what runs automatically, where AI helps and where a person decides.
Chase overdue invoices by email: once per customer, at the right level of firmness, never after payment.
View process map: Accounts Receivable Payment Chasing
Turn tender documents into reviewed, approved answers, with a person signing off at every point that matters.
View process map: Request for Proposal (RFP) & Tender Response Preparation System
Designed so a patient receives consultation access only after the payment is confirmed. Fictional clinic, test payments only.
View process map: Medical Clinic Website & Payment Integration
Turn a sales intake into a priced proposal that Sales reviews and approves before it reaches the client once.
View process map: AI Proposal Intake & Pricing System
Answer staff policy questions in the team chat, grounded in the company's own handbook documents. Tested with sample data.
View process map: RAG Employee Handbook Assistant
Answer customer questions about services on the business's Facebook page. Tested with sample data.
View process map: Messenger Inquiry Assistant
Photo of an invoice in, checked invoice record out. Tested with sample data.
View process map: Invoice Agent for Telegram
Answer sales questions in the team chat from the company's sales records.
View process map: SQL Sales Analytics Assistant
An AI voice assistant answers calls and books, moves or cancels appointments in the clinic calendar.
View process map: Dental Voice Appointment Setter
Knowledge-grounding experiment 1 of 3: the reference material is written into the assistant's instructions.
View process map: Full-Context Assistant: document in the prompt
Knowledge-grounding experiment 2 of 3: both reference documents are fetched in full for every message.
View process map: Full-Context Assistant: documents fetched per message
Knowledge-grounding experiment 3 of 3: the assistant decides for itself when to read the reference documents.
View process map: Full-Context Assistant: documents as tools
The first sales assistant, answering from a small sample sales table. Replaced by the SQL Sales Analytics Assistant.
View process map: AI Sales Data Assistant
When the team moves a lead to a new stage, the matching follow-up happens automatically.
View process map: Zapier–Asana CRM Automation
Enrich and score every new enquiry, then follow up according to how promising the lead is.
View process map: Zapier Lead Qualification & Outreach
Turn one recorded audio file into blog and social posts, publishing or holding them for review.
View process map: Zapier Content Repurposing
Move a car-insurance lead from first contact to an issued policy, with follow-up at each stage.
View process map: Car Insurance Sales Workflows
Take a salon client from first enquiry through their service to a rebooking follow-up.
View process map: Salon Client Retention Pipeline
Keep in touch with a new lead until they book an appointment, then welcome them as a client once paid.
View process map: Lead Nurturing Pipeline19 process maps
Stack
The tools behind the projects on this page. Every one was used in at least one of them.
How I build
I plan before I build: map how the work really happens, decide what should be automated and what stays with people, and look for what could go wrong — then build and test.
Write down how the work happens today, who does what, where decisions are made and where automation should stop.
Choose what fixed rules should handle, where AI is genuinely useful and where the trusted record of the data lives.
Imagine the system has failed and work out why. Identify weak points early and address them in the plan before building.
Agree the steps, data, tests and scope. A plan that isn't ready to build goes back for another round.
Test the normal path and the problem cases, such as blocked, duplicate and failed runs. Design checks and proof from real runs are reported separately.
Write down what is untested, what was deferred and the known limitations.
Deliberate failure testing is planned for each project but so far has only been partly carried out.
About
I'm Angelo Torrevillas, an AI Automation Specialist based in the Philippines. I build AI and automation systems that reduce repetitive business work. Connecting apps is the easy part. The real work is making the system dependable, so it keeps a clear record of what happened, follows fixed rules where accuracy matters, and asks a person to review the decisions that need one.
My background is in computer engineering. On every project I write down what has and hasn't been tested, because a demo should never imply more than it proves.
I take on client projects, and I'm also open to full-time, part-time, contract and freelance roles, ideally remote.
A camera-based device that measures a rebar column from a photo and recommends the concrete mix before pouring. It stayed within ±5% tolerance in 77 of 80 system tests (96.3%).
A web app built during my internship to replace manual attendance records at the Quezon City Public Library.
FAQ
He's an AI Automation Specialist who builds AI and automation systems that reduce repetitive business work, such as following up overdue invoices, answering routine customer questions, preparing proposals and keeping CRM records up to date, using the tools a team already has.
Tasks that follow the same steps and happen often, like reminders and follow-ups, copying details from documents into a system, answering common questions from your own documents, preparing proposals and updating a CRM. Tasks that need judgement can often be partly automated, with a person approving the final step.
Usually not. Most automations connect the tools you already have, such as email, spreadsheets, a CRM, chat apps and payment providers. We can check yours on a discovery call.
I use AI for work involving language, such as sorting replies, reading documents or drafting answers. Anything that must be exact, like payment status, follows fixed rules. When AI recommends something that affects money, such as a proposal fee, a person reviews it first. Chat assistants reply on their own, so they're built to answer from your documents, and their answers can still be wrong.
I start by mapping how the work really happens and looking for what could go wrong. The larger systems here include checks like re-checking payment just before a reminder is sent, blocking duplicate emails, a switch that stops sending at once, and approval steps for decisions like sending a proposal. Each case study says what was tested.
One is. The AI Proposal Intake & Pricing System was delivered during a paid trial for an interior design firm, which isn't named. The rest are portfolio systems built with sample data, test payments or a fictional business, and each shows its status.
Yes. The SQL Sales Analytics Assistant has a public demo on a sample sales dataset, and the medical clinic website has a public preview that uses test payments only.
Mainly n8n for workflows, PostgreSQL and Supabase for data, and GoHighLevel for CRM work. AI models come through OpenRouter, with Ollama for document search. Payments have been built with PayMongo in test mode, and messaging with Gmail, Slack, Telegram and Facebook Messenger. Earlier projects used Zapier, Airtable and Vapi.
With a free 30-minute Google Meet call about how your process runs today and where manual work is slowing it down.
Yes. Alongside client projects, I'm open to full-time, part-time, contract and freelance work, including short projects, and I prefer remote roles. My Upwork, JobStreet and OnlineJobs.ph profiles are linked below.
Book a call on this page, or email angelo.automates@gmail.com with a few lines about the process you want to automate.
Discovery call
A free 30-minute Google Meet call about how your process runs today, where manual work is slowing it down, and whether automation is the right fix.
How work moves today: the tools, the people and the hand-offs.
Where repetitive work, errors or delays are costing you time.
What to automate first, and what it takes to make it dependable.
Pick a time that suits you. Free 30-minute Google Meet call. The calendar loads as you scroll to it.
Calendar not loading? Open it in a new tab or email angelo.automates@gmail.com.
Contact
Prefer to write first? Send a few lines about the process you want to automate, or about the role you're hiring for. I'll tell you honestly whether automation is the right fix and what it would take to make it dependable.
Selected system · Accounts receivable
Concept Demo · FINAL means the four main workflows are complete and run on sample invoices. Connecting a real accounting system, an approval step for final notices, opt-out handling, alerts for replies that need a person, and monitoring are planned for a later phase.
Follows up overdue invoices by email, with one message per customer, firmer as the invoice ages, and a final payment check just before each email is sent.
Problem
Chasing overdue invoices is repetitive, but mistakes are expensive. A reminder sent after payment, three emails to one customer or a message to the wrong address damages the relationship. Simple automations can't rule those out.
Solution
Four connected workflows share one database that tracks every invoice and reminder. Fixed rules decide who to remind and how firmly. Each email is checked again just before it goes out, so a reminder for an invoice that has since been paid is cancelled. AI helps sort customer replies, fixed rules then pause or stop reminders, and the AI cannot change any financial record.
How it flows
Runs automatically
Where humans decide
Guardrails
UPDATE … RETURNING)Evidence · 4 images
Screenshots of the n8n editor canvas. Sticky notes are hidden; node names and connections are as built.
Stack: n8n · PostgreSQL · Gmail API · OpenRouter LLM (classification only) · JavaScript
Selected system · Bids & proposals
FINAL · structural completeness means every workflow is built and passes validation, and most were tested one at a time. The full system has not yet run end to end. Two smaller pieces are deferred, a document-length check in the review package and the hand-off from confirmed addenda to drafting.
Helps a bid team answer tenders by pulling out the requirements and drafting answers from previously approved content, with a person approving each answer and signing off the bid.
Problem
Answering a tender means pulling every requirement out of long, inconsistent documents, finding the best previously approved answer, drafting the rest and getting sign-off before a hard deadline. A missed requirement or an unapproved claim can cost the bid.
Solution
Several connected workflows share one database. They read the tender documents, pull out the requirements for a person to confirm, find the best previously approved answers and draft the rest with AI. Every answer then goes through review and approval, and only the bid owner can give final sign-off.
How it flows
Runs automatically
Where humans decide
Guardrails
Evidence · 8 images
Screenshots of the n8n editor canvas. Sticky notes are hidden; node names and connections are as built.
Stack: n8n · PostgreSQL + pgvector · Ollama embeddings · OpenRouter (GPT-4o mini) · Google Drive & Docs · n8n Forms with user auth · Slack
Selected system · Clinic website, booking & payments
A website and booking journey for a fictional clinic, designed so patients receive their online consultation link only after the payment provider confirms the payment and the clinic's system checks that the confirmation is genuine. Payments run in test mode only.
Visit demo site (opens in a new tab)Public GoHighLevel preview of the fictional clinic. Its checkout runs in PayMongo test mode only.
Problem
For paid online consultations, the consultation link is the product. A clinic needs patients to move smoothly from interest to booking, but access should only be released once payment is genuinely confirmed — not because a redirect page said so.
Solution
Built in two stages. First, the clinic's website pages and CRM workflows in GoHighLevel, which move patients through booking, send reminders and alert staff. Then a payment system in n8n with a database, which sets the price on the server, sends the patient to PayMongo's secure checkout, checks each payment confirmation and only then sends the consultation links and updates the CRM.
System layers and their status
Sandbox means PayMongo test mode — no real money moves. Verified once · controlled means one deliberate, supervised run has verified the CRM sync, which stays locked outside such runs.
Runs automatically
Where humans decide
Guardrails
Evidence · 22 images
Patient journey · GoHighLevel funnel to PayMongo sandbox · clinic name redacted in screenshots
CRM & booking · GoHighLevel workflows
Payment layer · n8n + PostgreSQL
Screenshots of the n8n editor canvas. Sticky notes are hidden; node names and connections are as built.
Stack: GoHighLevel · n8n · PostgreSQL (Supabase) · PayMongo · Webhooks + HMAC · Gmail API · HTML/CSS/JS
Client work · Paid trial
During the trial, the main path from intake to client email completed an end-to-end test, and the AI, PDF and email failure paths were each tested. Some safeguards aren't finished. There's no automatic check yet that keeps the AI's recommended fee within a set range of the rule-based fee, so Sales review is the safeguard. Proposals that need a pricing or manager review stop there, Sales can't yet adjust the price or request changes, and an invalid intake isn't recorded.
Turns a sales enquiry into a priced proposal PDF and makes sure only an approved proposal reaches the client, and only once.
Problem
The firm wanted design-fee proposals produced from sales intake under firm rules. The same inputs should give the same base fee, Sales must approve before anything is sent, and retries or edits must never duplicate a proposal or a client email.
Solution
One n8n workflow with more than 50 steps. Fixed pricing rules written in code calculate the base fee. AI then reviews the proposal for risks and missing information and recommends the final fee, which can include extra for complexity. Sales sees the rule-based fee and the AI recommendation side by side and must choose to continue before the proposal is built. The proposal is produced as a PDF for Sales to preview and approve, then emailed to the client with Sales copied.
How it flows
Runs automatically
Where humans decide
Guardrails
Evidence · 12 images
Workflow structure only — client name, pricing rules, proposal content and delivery evidence are withheld.
Stack: n8n Forms & Code · OpenRouter (Llama 3.3 70B) · Google Docs & Drive · Google Sheets records · Gmail · Slack error alerts
Build · n8n · AI agent
Lets a team ask questions about their sales data in plain English, such as monthly revenue or the top region, and get an answer with the figures. The original version is an n8n workflow used from Slack. A separate public web demo lets anyone try the idea on a sample dataset.
Visit Live Demo (opens in a new tab)A separate web demo, not the n8n and Slack workflow. It uses a sample sales dataset, and AI answers can be wrong.
The original workflow
Built in n8n and used from Slack. An AI agent looks up a sales table in PostgreSQL and uses a calculator before replying. It was tested on a sample dataset of 240 sales.
The public web demo
A separate small web app built for this portfolio, not the n8n workflow. AI writes one database query, a safety check allows only read-only queries on the sample table, and AI then summarises the result. It never touches a real database.
Evidence · 2 images
AI answers can be wrong. In testing, some answers from the Slack version contained incorrect dates or totals, so check important figures against the source data.
Stack: Slack · n8n · OpenRouter (GPT-4o mini) · PostgreSQL · public demo on a Vercel serverless function with PGlite and a SQL safety check
Build · n8n · Voice agent
An AI phone assistant for a dental clinic. During a call it uses five small n8n workflows to find free times and book, move or cancel appointments in Google Calendar, and logs bookings and call results in Airtable. Tested with test calls only. Built, but not currently running.
Evidence · 5 images
Stack: Vapi · n8n webhooks · Google Calendar · Airtable
Build · n8n · RAG
Answers staff policy questions in Slack using the company's own documents. New files added to a Google Drive folder are indexed for search (embedded with Ollama into a Supabase vector store). When someone asks, the assistant finds the relevant passages and answers from them. Built and tested with a sample handbook.
Evidence · 2 images
Stack: Google Drive · Ollama embeddings · Supabase vector store · OpenRouter · Slack
Build · Knowledge-grounding experiment 1 of 3
Tests the simplest way to give a chat assistant reference material, by writing it straight into its instructions. One of three small experiments comparing ways to ground answers in documents. The RAG assistant is the later approach.
Evidence · 2 images
Stack: n8n chat trigger · OpenRouter (GPT-4o mini)
Build · Knowledge-grounding experiment 2 of 3
Tests loading two full reference documents on every message, which gives complete context at a higher cost per message.
Evidence · 2 images
Stack: n8n · Google Docs · OpenRouter (GPT-4o mini)
Build · Knowledge-grounding experiment 3 of 3
Tests letting the assistant decide for itself when to open the reference documents.
Evidence · 2 images
Stack: n8n · Google Docs tools · OpenRouter (GPT-4o mini)
Build · n8n · AI agent
Answers customer questions sent to a business's Facebook page, using a Google Docs document that describes its services. It completes Facebook's connection check, ignores messages that aren't text, remembers the last few messages of a conversation and replies through Facebook's API. Built and tested with test data, and not used by real customers.
Evidence · 1 image
Stack: Messenger webhook · Google Docs · OpenRouter (GPT-4o mini) · Graph API
Build · n8n · AI agent
The first sales assistant: a chat agent that answers questions through four n8n Data Table tools — all rows, by product, by date, by transaction — plus a calculator. The SQL Sales Analytics Assistant replaced the Data Table tools with a PostgreSQL query tool.
Evidence · 2 images
Stack: n8n Data Tables · OpenRouter (GPT-4o mini)
Build · n8n · Document AI
Send a photo of an invoice to a Telegram bot. The text is read from the image (OCR), AI turns it into invoice fields, and missing details or duplicate invoices are reported back before anything is saved. The image goes to Google Drive, the record to Google Sheets, and a confirmation comes back in the chat. Built and tested with test invoices. A separate error workflow is connected to report failures by Telegram. It runs automatically, but a delivered alert hasn't been confirmed yet.
Evidence · 1 image
Stack: Telegram · OCR.space · OpenRouter (GPT-4o mini) · Google Drive · Google Sheets
Earlier build · Zapier
An updated Asana task splits into five paths — Ready to Start, No Response, Quoted, Approved, and Paid and Closed — that create lead folders and sub-tasks, send timed Gmail follow-ups, and draft welcome and recommendation emails with AI.
Evidence · 1 image
Stack: Zapier · Asana · Google Drive · Gmail · AI by Zapier
Earlier build · Zapier
A form submission is enriched with company data from Apollo, scored by AI and saved to a Supabase database, then routed by priority: high and medium leads are automatically sent an AI-written outreach or nurturing email, with a Slack alert; low leads get a summary in Slack.
Evidence · 1 image
Stack: Zapier · Apollo API · AI by Zapier · Supabase REST · Gmail · Slack
Earlier build · Zapier
An audio file added to Google Drive is transcribed and turned into two blog drafts and Facebook and LinkedIn posts. An AI check then posts them to my own Facebook and LinkedIn accounts or saves them to Google Docs for a person to review, and every stage is logged in Google Sheets.
Evidence · 1 image
Stack: Zapier · Google Drive · AI by Zapier · Google Sheets · Google Docs · Facebook Pages · LinkedIn
Earlier build · GoHighLevel
A set of GoHighLevel workflows for a car-insurance sales process. The first takes a new or tagged contact, creates an opportunity, tags it, sends the insurance intake form, assigns an available agent and notifies the team.
Evidence · 5 images
Stack: GoHighLevel workflows · email
Earlier build · GoHighLevel
GoHighLevel workflows for a salon. New-lead intake turns an inquiry form into an opportunity with a tag, a welcome email, a front-desk notification, a note and a “contact within 1 hour” task.
Evidence · 4 images
Stack: GoHighLevel workflows · email
Earlier build · GoHighLevel
GoHighLevel workflows that move a lead through tagged stages. When the appointment form is submitted, the “new lead” and “nurturing” tags are removed and “booked” is added.
Evidence · 4 images
Stack: GoHighLevel workflows · forms
Background · Team thesis
A Raspberry Pi 5 with a camera and distance sensor photographs a rebar column; a Mask R-CNN model — chosen after evaluating YOLOv8 and RetinaNet — detects it, the system computes its dimensions and concrete volume, and a web app recommends the cement mixture.
My role
I started as the deep-learning model developer, was reassigned to web-application development in the final phase, and integrated the move from a laptop-hosted web app to a Raspberry Pi deployment.
Trials covered rebar-column detection (19/20), non-rebar detection (20/20), dimension and volume computation (19/20) and cement-mixture estimation (19/20).
Screens · 6 images
Stack: Python · Mask R-CNN · OpenCV · Flask · Raspberry Pi 5
Background · Internship · Quezon City Public Library
A web application built during my internship to replace manual attendance records, with time-in and time-out, attendance history, an admin dashboard and reports.
Screens · 5 images
Stack: PHP · MySQL · JavaScript · HTML/CSS
Process map · Accounts receivable
Chase overdue invoices by email: once per customer, at the right level of firmness, never after payment.
The challenge
Chasing overdue invoices is repetitive, but mistakes are costly: a reminder after payment, several emails to one customer or a message to the wrong address damages the relationship.
The approach
Fixed rules decide who to chase and how firmly, every reminder is re-checked just before it's sent, and customer replies are sorted by an AI classifier that can't change any financial records.
Process map · Bids & proposals
Turn tender documents into reviewed, approved answers, with a person signing off at every point that matters.
The challenge
Answering a tender means pulling every requirement out of long, inconsistent documents and getting approved answers signed off before a hard deadline. A missed requirement or an unapproved claim can cost the bid.
The approach
Requirements are extracted and confirmed by a person, answers are drafted only from an approved library, and every answer goes through review, approval and the owner's final sign-off.
Process map · Clinic website, booking & payments
Designed so a patient receives consultation access only after the payment is confirmed. Fictional clinic, test payments only.
The challenge
For paid online consultations, the consultation link is the product: access should only be released once payment is genuinely confirmed, not because a confirmation page said so.
The approach
A patient journey from choosing a doctor to payment, where the clinic's system sets the price, checks that the payment confirmation is genuine and only then sends the consultation links. A separate CRM process follows up on bookings and visits.
Process map · Interior design firm · Anonymized
Turn a sales intake into a priced proposal that Sales reviews and approves before it reaches the client once.
The challenge
The firm needed design-fee proposals from sales intake where the same inputs always give the same base fee, Sales approves before anything is sent, and retries never duplicate a proposal or a client email.
The approach
Rule-based pricing sets the base fee, AI reviews the risks and recommends the final fee, Sales reviews both and approves the PDF, and the client receives the proposal once.
Process map · n8n · RAG
Answer staff policy questions in the team chat, grounded in the company's own handbook documents. Tested with sample data.
The challenge
Staff have policy questions whose answers are in the company's handbook documents.
The approach
New handbook documents are indexed as they're added; when someone asks in Slack, AI finds the relevant passages and answers from them.
Process map · n8n · AI agent
Answer customer questions about services on the business's Facebook page. Tested with sample data.
The challenge
Customers send inquiries about the business's services through Facebook Messenger.
The approach
Text messages are answered by AI from a single services document, with short conversation memory; other message types are ignored.
Process map · n8n · Document AI
Photo of an invoice in, checked invoice record out. Tested with sample data.
The challenge
Invoice photos need to become structured records, with missing details and duplicate invoices caught before anything is stored.
The approach
The text is read from the photo, AI pulls out the invoice details, and missing details or duplicates are reported back before anything is stored.
Process map · n8n · AI agent
Answer sales questions in the team chat from the company's sales records.
The challenge
Sales questions need answers drawn from the sales records, including calculated totals.
The approach
Questions asked in Slack are answered by an AI agent that looks up the sales records and uses a calculator. Answer accuracy is still being evaluated.
Process map · n8n · Voice agent
An AI voice assistant answers calls and books, moves or cancels appointments in the clinic calendar.
The challenge
Callers need to check, book, move or cancel appointments over the phone.
The approach
An AI voice assistant handles the call and uses calendar actions to find free times and book, move or cancel appointments, recording the outcome of every call.
Process map · n8n · Experiment
Knowledge-grounding experiment 1 of 3: the reference material is written into the assistant's instructions.
The challenge
Comparing ways to give a chat assistant the reference material it answers from.
The approach
The reference material is written directly into the assistant's instructions, with no lookup step.
Process map · n8n · Experiment
Knowledge-grounding experiment 2 of 3: both reference documents are fetched in full for every message.
The challenge
Comparing ways to give a chat assistant the reference material it answers from.
The approach
Both reference documents are fetched in full on every message, trading cost per message for complete context.
Process map · n8n · Experiment
Knowledge-grounding experiment 3 of 3: the assistant decides for itself when to read the reference documents.
The challenge
Comparing ways to give a chat assistant the reference material it answers from.
The approach
The documents are available as tools, so the assistant decides for itself when to read them.
Process map · n8n · AI agent
The first sales assistant, answering from a small sample sales table. Replaced by the SQL Sales Analytics Assistant.
The challenge
An early test of answering sales questions from a small sample dataset.
The approach
A chat assistant looks up sales by product, date or transaction and calculates totals. It was later replaced by the SQL Sales Analytics Assistant.
Process map · Zapier · CRM
When the team moves a lead to a new stage, the matching follow-up happens automatically.
The challenge
Each stage a lead reaches on the project board calls for a different follow-up.
The approach
When the team updates a lead's stage, the matching follow-up runs: a project folder and task, timed follow-up emails, quote follow-ups, or an AI-written welcome or recommendation email.
Process map · Zapier · Lead qualification
Enrich and score every new enquiry, then follow up according to how promising the lead is.
The challenge
New form enquiries need enriching with company data and prioritising, so each gets the right follow-up.
The approach
Each enquiry is enriched with company details and scored by AI; high and medium leads are automatically sent an AI-written email with a sales alert, and low leads are summarised for sales.
Process map · Zapier · Content
Turn one recorded audio file into blog and social posts, publishing or holding them for review.
The challenge
One audio recording needs to become blog posts and social posts.
The approach
The audio is transcribed and drafted into blog and social posts by AI; an AI check then publishes them or saves them for a person to review.
Process map · GoHighLevel · CRM
Move a car-insurance lead from first contact to an issued policy, with follow-up at each stage.
The challenge
A car-insurance lead passes through several stages, from first contact to an issued policy, each with its own follow-up.
The approach
Each stage triggers its own emails, assignments and team alerts, from the intake form and package options to upsell offers and policy delivery.
Process map · GoHighLevel · CRM
Take a salon client from first enquiry through their service to a rebooking follow-up.
The challenge
A salon client moves from first enquiry, through their service, to rebooking, and each stage has its own follow-up.
The approach
The enquiry creates the client record and a welcome email; as staff move the client through each stage, the matching emails, notes and tasks follow, ending with a rebooking offer.
Process map · GoHighLevel · CRM
Keep in touch with a new lead until they book an appointment, then welcome them as a client once paid.
The challenge
New leads who don't book straight away need following up until they do.
The approach
Leads get a booking link, then a nurturing email each day until they book; the booking is confirmed, and paid clients are moved to the client stage and thanked.
