AboutServicesWorkCVContact

Most of the work draining your team can run itself.

John Lemuel Culinares

John Culinares

AI automation specialist · Philippines

Open to new projectsContact me

The stack, as one thing

Based inPhilippines, on UK and US hours
Clients inthe UK, the US and Australia
Who I build forAgencies, property and real estate
How I workOne person, first call to handover

About me

I build AI systems for businesses in the UK, the US and Australia, from a desk in the Philippines. Most projects start as a spreadsheet somebody updates by hand, a call nobody answers, or a report that takes a day to assemble. I build the thing that ends it.

Contact me

Services

One person from first call to handover, so nothing gets lost in a hand-off.

01

Voice agents & automation

Voice agents that answer and qualify calls around the clock, and workflows that carry a lead from first contact to booked appointment without anyone retyping it.

Voice AI receptionists on Retell · n8n and Zapier pipelines · Missed-call text-back and follow-up
02

Internal tools & portals

Client portals, internal dashboards and the small apps that remove a spreadsheet from your process. Built to be handed over, documented, and run without me.

Client portals and internal tools · Next.js, Python, Supabase · Tested and documented on handover
03

Connected systems

Two-way syncs between the tools you already pay for, with the boring parts handled: duplicate matching, conflict holds, retries, and an alert when something genuinely breaks.

CRM, accounting and ad platforms · Two-way sync with conflict handling · Webhooks, REST APIs, MCP servers
04

Reporting that decides

Numbers pulled from every platform into one place, then an AI layer on top that returns a decision instead of another chart nobody opens.

Multi-platform ad and SEO reporting · Scale, kill or iterate calls per ad · Daily batch jobs and Slack digests

Selected work

Live systems in use today. On client systems, names are withheld.

01

Client project · Portal and CRM integration

Housing Portal + Zoho CRM Sync

650+ automated tests · 264 records in, 252 out

A client portal bolted onto a property group's existing site, kept in step with Zoho CRM in both directions, so the same property stops being typed twice.

  • Owners submit, providers shortlistProperty owners put a house in, care providers filter by beds and area and request a viewing, all without an email.
  • Two way, not two copiesEdit in the portal or edit in Zoho. Both land, and neither overwrites the other with stale data.
  • Address matching that caught the duplicatesNormalising addresses before comparing them found 12 duplicate pairs and took 264 records down to 252.
02

Client project · Reporting warehouse

Prism Marketing Dashboard

187 accounts · 5 ad platforms + 3 SEO sources · 25 views

Five ad platforms and three SEO sources pulled into one warehouse, with 25 analysis views, budget pacing, wasted-spend detection and plain English querying over the top.

  • Eight sources, one tableFive ad platforms and three SEO sources land in the same schema every night.
  • Budget pacingShows where an account will land by month end against what it was given, early enough to act.
  • Wasted spend detectionFlags the campaigns and keywords taking budget without returning anything.
03

Client project · Voice AI

Grace Inbound AI Receptionist

24/7 coverage · three agents live

Three voice agents on real calls. One takes inbound seller calls, one qualifies inbound real estate leads, one rings out and pre-qualifies with SMS follow-up. All three hand a clean record to the CRM.

  • Answers every callNo voicemail, no office hours, no queue. A seller ringing at 2am gets the same qualifying conversation.
  • Qualifies, not just answersAsks about the parcel, the situation and the timeline, and knows which answers end the call politely.
  • Outbound with SMS follow-upSARAH rings out and texts the ones who do not pick up, so a lead is not dropped after one attempt.
04

Own product · Open source

Google Ads MCP Server

An open-source MCP server that lets Claude or ChatGPT query and manage a Google Ads account natively, in conversation, instead of through an export.

  • Campaign and keyword reportingSpend, clicks, conversions and cost per conversion over any date range, without writing GAQL.
  • Flat rows, not nested objectsResults come back shaped for reading, so the model answers about the field you meant.
  • Search termsWhat people actually typed, which is where the wasted spend hides.
05

Client project · Attribution

Pipedrive to Google Ads

An n8n workflow that maps Pipedrive deal stages to Google Ads offline conversion events, so the ad platform optimises on deals closed rather than forms filled.

  • Stage changes become conversionsQualified, proposal and won each upload as their own conversion action.
  • Values, not just countsDeal value goes up with the conversion, so bidding optimises on revenue rather than volume.
  • Click ID carried throughThe GCLID is captured at the form and stored on the deal, which is what makes the match possible.
06

Client project · AI evaluation

Copy Chief AI

7 scoring criteria

An evaluator that scores marketing copy against a seven-criterion rubric and says why, so feedback stops depending on who read it that day.

  • Seven criteria, scored apartHook, clarity, offer and the rest are judged individually, so a strong hook cannot carry a weak offer.
  • Reasons tied to linesEach score quotes the copy it is judging, not the piece in general.
  • Anchored scalesEach criterion defines what a 2 and a 9 look like, which is what stops everything landing on 7.
The build logEverything else28 projects, five dashboards you can click around in, and a recorded call from one of the voice agents.

How a project goes

Clear scope and price before any work starts.

01

Find the bottleneck

A call, then I go looking at where the time actually goes. Usually it is not the thing you think it is.

02

Build the smallest version that works

One workflow, one agent, one screen. Something you can judge in a week rather than a quarter.

03

Test it against real data

Edge cases, bad input, the call that goes sideways. If it cannot survive your worst Tuesday it is not finished.

04

Hand it over documented

SOPs and a walkthrough so your team runs it. When the project ends you are not tied to me.

What I build with

Agents and models

Claude API, OpenAI, MCP servers, custom GPTs

Voice

Retell AI, Vapi, Twilio SMS

Automation

n8n, Make, Zapier, Google Apps Script

Apps

Next.js, React, TypeScript, tRPC, Python

Data

Supabase, PostgreSQL, MySQL, BigQuery, Sheets

Platforms

GoHighLevel, Zoho CRM, Pipedrive, Google Ads, Meta

Need a system built?

Or tell me here

Goes to me only. No list, no follow-up sequence.

AskJohn