Raza Rizwan · Forward Deployed Engineer
ChatGPT and Claude did not remove the human from software development. They moved the human to the keyboard. Someone still types every instruction, re-explains the project after every context reset, and cleans up the rework.
I put your delivery process inside the AI as an MCP connector, so it researches, designs, builds, tests and deploys the same way every run, and stops for a person only at the approval gate.
Forward deployed means I work inside your real workflow, connect the tools your team already uses, and stay accountable for what ships.
One workflow first · Success criteria agreed upfront · Human approval where it matters
8live MCP connectors, checked today
399automated checks across 7 connectors
6years building AI systems
111GitHub repositories
REALITY CHECK / WHAT CHATGPT DID NOT AUTOMATE
Look at any "AI-built" project honestly and a human is still doing four jobs. Each one is exactly where my systems go.
PROOF / PROBLEM → ACTION → RESULT
No adjectives. What was broken, what I did, what changed. A number in every result.
Every AI-built landing page needed a person to re-type the brief, the design rules and the QA checklist, every session. Pages shipped with placeholder copy and no mobile check.
I converted a 44-skill full-stack agent into an MCP connector whose server rejects out-of-order work: 7 design docs before code, a skill read proven per file, placeholder copy rejected on write, a 12-point hero audit, and gates that reset on every edit.
Brief → Design → Build → QA → Delivery now runs identically from ChatGPT, Claude Code or Codex. 42 smoke + 16 live checks green. 7 of 10 steps need no human.
Each connector was a from-scratch build, and the formula lived in my head: the same context loss I fix for clients.
I wrote the formula down as an executable skill: a 10-step path, 8 references, 8 templates and a scaffolder that generates a working server with the gates already marked.
A fresh scaffold passes 18/18 smoke checks with zero edits. Later connectors went from skills to a live URL in one day (canz-ads, canz-pm). 8 connectors live today.
Seven people, ~3,600 board tasks, and the real "how" scattered across a database and repos. Every new hire, and every AI session, relearned it from zero.
I extracted 5 SOPs into 1,784 prompt files (verbatim, SHA-256 verified) and packaged them into 4 AI agents (research, website, marketing, design) plus a QA judge with a 12-dimension rubric. Humans approve at one Quality Gate.
Deployed 23 July 2026 on two machines, 3/3 smoke passed, 0% change to the existing pipeline. Every deliverable now scores /100; nothing launches under 80.
DIRECTION / WHERE ONE WORKFLOW LEADS
Work enters from the CRM. AI staff do the job. One person reviews. The client receives it. Nobody types a brief, chases a status, or assembles a handover.
THE HUMAN'S DAY
COUNTED
The agency in story 03 runs this way today. How it is built is not published. Ask me in the audit ↗
SELECTED DEPLOYMENTS / PUBLICLY VERIFIABLE
Two of the eight live connectors, shown in full. Each one turns a job into a defined, reviewable delivery process. Measured client outcomes appear only when verified.
01 / LIVE MCP SERVER
Turns a landing-page request into a design-first, quality-gated delivery process. 44 skills, 7 design docs, hero 12/12, secrets and placeholders unwaivable.
02 / LIVE MCP SERVER
Turns a rough development request into a researched plan, tested changes, and a reviewable delivery package. The first code file must be a test.
BUSINESS WORKFLOWS
The technology comes second. First, define the completed job, the evidence it needs, and where a person must remain in control.
Brief, design, build, QA and deploy as one gated run: the AI executes, the server enforces, a person approves twice.
Research a request, plan the steps, use approved tools, test the work, and prepare it for review.
Turn intake into a structured proposal, an exact-quantity ad set, or a daily board report, with human review before delivery.
SERVICES / CLEAR STARTING POINTS
The scope gets wider only after one workflow and its success criteria are clear.
Map one process step by step, count where the humans sit today, find the highest-value automation point, and recommend a focused pilot.
Start with an audit ↗Build one limited workflow with required integrations, approvals, logging, evaluation, and handover.
Discuss a pilot ↗Connect live tools, set permissions, test failure paths, document the system, and prepare the team.
Plan a deployment ↗Review failures, improve prompts and evaluations, and add new uses when the system proves useful.
Discuss optimization ↗METHOD / FOUR CONTROLLED MOVES
Each stage earns the next. The system touches more of the operation only when the evidence supports it.
RELIABILITY / HUMAN CONTROL
The 30% that stays human is the part that should: the brief, the design decision, the go-live. Everything else is enforced, logged and reviewable.
HIGH-IMPACT ACTION
WAITING FOR HUMANReview evidence · Approve · Return for revision
JOURNEY / 2020 TO 2026
Follow the path from predictive AI and MLOps to agents, AI employees, and forward deployed engineering.
Read the journey, year by year ↗SOURCE / GITHUB
Browse 111 repositories across the six-year journey.
Open GitHub ↗BUYER FAQ / DIRECT ANSWERS
It is counted per run, not guessed. A landing-page run through the website connector has 10 steps; 3 are human decisions (brief, design approval, ship approval) and 7 are executed and checked by the system. The count for your workflow is agreed during the audit.
Nothing is lost. Every build has an ID and its state lives in a database, not in the chat. A new session resumes from the last completed step instead of re-explaining the project.
Yes. One already does. Work enters from the CRM, AI staff do the jobs, and one person reviews and approves. It started as a single workflow; the audit is where yours would start too.
We map one current process, count where the humans sit, identify the best automation point, define what success means, and decide whether a focused pilot is worth building.
The timeline depends on the workflow, integrations, approval paths, and evidence required. The audit defines a realistic scope before implementation starts.
That is the default. The connectors run from ChatGPT, Claude Code or Codex, and the workflow is designed around the tools, permissions, and operating constraints your team already uses.
High-impact actions keep a human gate. The exact approval points are agreed before the system is deployed.
Each system role receives only the access it needs. Data paths, permissions, and logs are reviewed as part of the workflow design.
Failure paths are designed before launch. The system can retry safely, stop for review, or return the work for correction.
The deployment includes documentation and handover. Ongoing optimization can review failures, improve evaluations, and extend proven workflows.
CONTACT / WORKFLOW AUDIT
Bring one workflow where a person still babysits the AI, or the whole operation. In the first conversation we map the run step by step, count where the humans sit, and decide whether a focused pilot makes sense.