I automate what companies do by hand.

In production right now: an invoice platform that extracts, validates and files 2,000+ financial documents a month. Data pipelines that have moved 100K+ records. A content engine that brought in qualified leads within 30 days. Four AI systems shipped, each one replacing work that used to be done by hand.

How I turn manual work into systems that run on their own.

Every project follows the same principles. They are the reason the systems keep running in production instead of becoming another abandoned pilot.

1. 1

   Map the manual work first

   Before any code, I sit with the process as it runs today: where the data arrives, who touches it, what they check, where it goes and where it breaks. Most automation projects fail because they automate an imagined process instead of the real one. A clear map also shows which steps are worth automating first, usually the ones that are frequent, repetitive and error-prone.

2. 2

   Design the pipeline, not just the model

   A useful AI system is mostly plumbing: ingestion from the places data already lives, extraction, categorization, validation, assignment and delivery into the systems people already use, such as Outlook, Odoo or PostgreSQL. The language model is one step inside that pipeline. When people say AI is just a chatbot with an LLM behind it, this is the part they are missing.

3. 3

   Put deterministic checks around every AI step

   Language models are flexible and occasionally wrong. Rules are rigid and predictable. Every generative step in my systems is wrapped in checks: schema validation, reconciliation of totals, required fields, forbidden claims. Anything that fails is flagged with a specific reason, so people only look at the cases that genuinely need judgement.

4. 4

   Keep a human where it counts

   The best place for a person is usually at the final decision, not in the data entry. My sales agent drafts replies instead of sending them; my invoice pipeline routes rejected documents to a review queue. People stay accountable, mistakes stay cheap, and trust grows with every draft that needs no edits.

5. 5

   Ship small, measure, then expand

   I prefer to put one well-defined flow into production and measure it (time to reply, documents processed, rejections, hours saved) before widening the scope. Real usage teaches more than any plan, and a working system earns the trust needed for the next one.

What I build.

Four kinds of systems, each one already running in production for real businesses.

• 22 September 2026

  Beyond the chatbot: an AI agent that ingests, categorizes and routes 100+ leads a day

  How an email agent on Microsoft Graph categorizes inbound leads, researches each company, drafts replies in Outlook and assigns leads round-robin by region and seniority, with PostgreSQL synced to Odoo.

• 15 September 2026

  The content engine that reached page one in two months

  Topic and keyword research, EEAT-aligned briefs, a deterministic quality gate and automated publishing: how an AI content platform brought qualified leads in 30 days.

• 8 September 2026

  Automating LHDN e-Invoice compliance for 2,000+ documents a month

  How a multi-tenant pipeline ingests invoices from mailboxes, extracts PDFs and spreadsheets, validates, transforms and submits them for Malaysian e-Invoicing.

• 1 September 2026

  Building a voice AI assistant that doesn't feel like a walkie-talkie

  Latency budgets, streaming, interruption handling and provider failover in a real-time voice assistant built with Pipecat, WebRTC, Whisper, LLaMA and ElevenLabs.