The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Building software with AI requires more than a good prompt.
I've been developing my own Autonomous Engineering Harness to structure how I build software with AI.
The idea is simple: instead of relying on AI to generate code one task at a time, I wanted a more structured process that connects specifications, task planning, implementation, code review, and QA.
The harness organizes the workflow into clear stages:
- Specifications: define what needs to be built.
- Task planning: break requirements into manageable units.
- Implementation: coordinate AI-assisted development.
- Review: check the result against the intended requirements.
-QA: validate the software before considering the work complete.
The goal isn't to make AI autonomous for the sake of it. It's to make AI-assisted engineering more structured, repeatable, and easier to validate.
I'm applying this approach to my own projects and continuing to refine the framework as I learn what works in practice.
I wrote more about the approach, the engineering decisions behind it, and how the framework works on my website.
Read the full story: https://yitsymosqueda.com/autonomous-engineering
I'd carry acceptance criteria from the spec into QA as explicit checks, with a result attached to each one. When the spec changes mid-build, does the workflow flag which completed tasks and QA results need to be revisited?
In my Autonomous Engineering Harness, the Product Agent generates tickets from specs, including explicit Acceptance Criteria and a Definition of Done. Code Review and QA use these to validate the implementation against the ticket's requirements.
Specs aren't changed mid-build....
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
AI Automation Platform and Workflow Automation & AI-Powered Solutions
An AI-powered automation platform designed to simplify business processes, automate repetitive tasks, and improve overall productivity.
The platform combines a modern, user-friendly interface with intelligent automation to help businesses streamline workflows, reduce manual effort, and work more efficiently.
The motion, the UI, everything's amazing, Incredible work
AI isn't the designer. It's the tool.
You're the designer. The art director.
It can't design without your direction.
yes, we are like a boss to AI. its all upto us how we get done our work from subordinate.
I'm a medical doctor. The scariest decisions in medicine happen where there is no doctor.
When someone falls sick far from a hospital, a health worker (or the family) has to decide in seconds: is this an emergency, or can it wait? Guess wrong, and someone can die waiting.
For the Base44 Challenge I didn't want to ship a pretty landing page. I built the tool I wish those health workers had.
Meet TriageLine: WHO-grounded triage decision support, built entirely in Base44.
THE KEY IDEA
Most AI health tools let the model make things up. TriageLine does not. The AI only NAMES the condition. A verified WHO protocol table then decides the severity, the exact dose, and the next step, with a page citation on every card. The AI cannot change a number on it. And if no protocol matches, it abstains: no citation, no plan, refer to a clinician.
HOW IT WORKS
Type the case in plain words, one free-text box: "child, 2 years, cough, fast breathing, too weak to feed."
The AI classifies it against the WHO list.
A verified WHO table (IMCI and mhGAP) returns the severity band, the exact dose, the next step, and the citation.
If it is an EMERGENCY, TriageLine automatically emails a doctor the full clinical summary to come in and escalate, and gives the patient an action-only summary.
Every case is recorded: a triage queue, an audit log, and analytics including time saved.
WHAT I BUILT
One-box intake plus AI classification across 30+ conditions
A verified WHO protocol table that binds every dose and citation, so no numbers are ever invented
EMERGENCY email escalation to a doctor, a real working automation
A patient-facing, action-only summary
A staff console: triage queue, audit log, analytics, role references
A safety abstain path when no WHO protocol matches
A dark, clinical design with a signature Pathway decision-spine that draws the one-way flow: the AI classifies, the table decides
THE PROCESS
As a doctor, I went deep into the WHO triage protocols, the IMCI and mhGAP guidelines written for exactly these settings, and worked out how each case should map to a severity, a dose, and a next step. Then I designed the trust architecture (AI classifies, verified table decides) and built all of it in Base44: the data model, the roles and auth, the AI classification, the WHO protocol table, and the email escalation. Then I tested it live end to end, a real emergency case producing the cited card and a real delivered email to the doctor.
It is live. Try it: triage-line-flow.base44.app
#Base44Challenge @Base44
I'd store the model's classification and the protocol-table version with each case, so the audit log can reproduce why that card appeared. When you update a WHO table, can you replay a fixed set of synthetic cases and review the output changes before the new version goes live?