AI Development & Automation

AI applications, intelligent assistants, conversational AI, AI-powered workflows and custom business automation, with a human in the loop where it matters.

What I build

Scheduled agents
Jobs that run on their own timetable, like a daily agent that picks a theme from a content strategy and writes the post and image prompt.
Human approval
AI output lands in a dashboard to review, edit, approve or reject. Nothing goes out until a person says so.
Observable by design
Every agent step is written to a log, and jobs are safe to re-run, so a retry can't create a duplicate.
Self-hosted image generation
ComfyUI running on cloud GPUs, including tracing and fixing the infrastructure failures underneath it.

Concepts

Automation blueprints

Designs for common business problems, each with a person in control. These are concepts shown with sample data, not client deployments.

Blueprint 01
Help-centre assistant with human handoff
Customers ask the same questions all day, but some conversations need a person.
1. 1Answers from the help centre and shows which article it used.
2. 2Hands off when it isn't confident, the customer asks for a person, or money is involved.
3. 3Gives the agent a summary: the issue, what was tried and why it was escalated.
Person in control: Refunds above a set amount always need a person's sign-off.
- RAG
- OpenAI API
- Vector search
- Escalation rules
Support chatSample data
Can I return a ring after 10 days?
Yes. Unworn pieces can be returned within 30 days with the receipt.
Returns policy, section 2
Then refund my ₱18,500 today.
Refund over ₱10,000Confidence 41%
Handed to a support agent
Issue: refund request, ₱18,500
Tried: answered with the returns policy
Why: refund over the ₱10,000 limit
Blueprint 02
Invoice extraction with a review queue
Supplier invoices arrive as PDFs and get retyped into a spreadsheet by hand.
1. 1Reads each PDF into fields: supplier, invoice number, line items and totals.
2. 2Checks the numbers: items add up to the total and no field is missing.
3. 3Saves clean rows to Sheets and sends uncertain fields to a review queue.
Person in control: Nothing uncertain is saved silently. A person confirms every flagged field.
- Document extraction
- OpenAI API
- Validation rules
- Google Sheets
Blueprint 03
Lead scoring and follow-up agent
New inquiries wait hours for a reply, and a ready buyer looks the same as a browser.
1. 1Scores each inquiry on budget, timeline and fit.
2. 2Drafts a personal reply and posts it to Slack for approval.
3. 3Logs the lead in the CRM and schedules follow-ups until the customer replies.
Person in control: Every reply is approved or edited by a person before it's sent.
- n8n
- OpenAI API
- Slack
- CRM
Blueprint 04
Ticket triage and weekly customer insights
Tickets land in one inbox, and nobody has time to spot the patterns.
1. 1Tags each ticket by topic, urgency and sentiment.
2. 2Routes it to the right team's queue.
3. 3Writes a weekly summary of the top issues and how they're trending.
Person in control: Agents can fix a wrong tag in one click, and each fix is used to measure accuracy.
- Classification
- Sentiment analysis
- Routing
- Reporting
Blueprint 05
Testing and monitoring for AI workflows
AI features drift quietly: one prompt change can break answers nobody re-checks.
1. 1Runs each workflow against labelled test cases after every change.
2. 2Tracks accuracy, cost per run, response time and failures over time.
3. 3Flags regressions and shows the exact cases that failed.
Person in control: A change only ships when it passes the test set.
- Evals
- Test sets
- Logging
- Dashboards