Forward Deployed Engineer · Johannesburg, US hours

I go into a business, map how the work really flows, and build the system that runs it.

Then I stay until it is in daily use. Most implementations fail not because the technology is hard, but because nobody mapped how the work actually happens before they started building. That mapping is the job.

Members on file4,277
Emails delivered273
Human open rate42.7%
Follow-up gap closed33 months

The work

Implementations that are running, or that ran to completion. No mockups, no concept work.

Healthcare · United States · Completed September 2026

A follow-up system for a membership practice, from zero

A US healthcare practice had a large member base and no follow-up system. I mapped how enquiries actually moved through the business, then built the command centre, the CRM and email integration, and a campaign engine with membership exclusions and a human approval gate on every single send.

The engagement ran sixteen months and completed in September 2026. It was handed over with functional specs, data models, runbooks and leadership reporting, and the client team ran it without me.

4,277Members on file
273Emails delivered
42.7%Human open rate
33 moFollow-up gap closed
Construction · Riyadh · Live

A commissioning dashboard, hosted and working the same day

Commissioning progress across HVAC, firefighting and plumbing disciplines at KAFD, Riyadh. The problem was that progress lived in a dozen disconnected places and nobody could see the whole picture. Built and hosted within a day of the first conversation, and still standing.

Hospitality · Live

A concierge that answers at 2am

A lodge was losing bookings because nobody answered WhatsApp overnight. The concierge handles WhatsApp, voice and web enquiries around the clock, pulls rates from the property's own site, and escalates to staff when a human should take over. It was tested on the owner's phone before the explanation finished.

Real estate · Live

A voice agent for the calls the office misses

An estate agency website and marketing system, with an AI voice agent that answers missed calls, qualifies the caller and books the viewing. The missed call is where the instruction is lost, so that is where the system sits.

Dental · South Africa · Live

Night-time enquiries, answered in each practice's voice

Lead-capture chatbots for three dental practices, built so the system never touches clinical judgement. It handles the enquiry, the intake and the follow-up, and hands to a human for anything clinical.

The Flow

The pipeline pattern behind all of it, written as a working system you can run yourself. Nine stages, twenty tests, and a deliberate failure script.

  1. Source records arrive and are logged with an id system
  2. Intake batch selection, in fixed sizes system
  3. Validate required fields, exclusions, duplicates. No model system
  4. Route clean records proceed, everything else is reason-coded system
  5. Agent step drafts, researches, builds agent
  6. Approval gate a named human approves. The agent can block, never send human
  7. Action the send, the write, the update system
  8. Receipt stage, actor, timestamp, request id — never the content system
  9. Measurement computed from the receipts, not typed by hand system

It runs on a real reviewed batch of 75 records and reproduces the reviewer's actual decisions, 69 approved and 6 rejected. The receipts are hash-chained, so a tampered log fails verification. There is a script that breaks the pipeline three ways on purpose, because a system you cannot break on demand is a system you do not understand.

How I work

The discipline is the product. These are the rules I hold on every build.

Rule 01

Deterministic before agent

If a rule can decide it, it is code, not a prompt. Models are used where judgement is genuinely needed, and nowhere else.

Rule 02

A human approves the send

The system can block. It can never send. Every outbound action passes a named human, and the machine enforces the difference between a hard block and a flag.

Rule 03

The receipt logs the doing, not the content

Stage, actor, timestamp, model, request id. Never what was said. A full audit trail that does not become a second copy of the data.

Rule 04

Go-live is not the finish line

The outcome baseline is agreed with the owner before the build starts, and definition of done is agreed together. Adoption is designed in, not hoped for.

About

I work in the room with the customer, not from a ticket queue. Discovery, scoping, the build, the go-live, and the awkward week after it.

Twenty years in sales and marketing before this, which is why I am comfortable in front of a customer and why I can tell a technical story to someone who is not technical.

Johannesburg, working US hours. Remote roles and fixed-scope engagements.

Contact

Remote roles and fixed-scope engagements. I work US hours from Johannesburg.

Email

john@aithatbooks.co.za

LinkedIn

linkedin.com/in/john-bianchina