Overview
A multi-agent software engineering system that takes a software requirement in plain English and drives it through the full SDLC lifecycle using a team of specialized AI agents: requirements analysis, design, planning, implementation, testing, and documentation. Between each stage the system pauses for human approval, where a reviewer can approve the output or reject it with feedback that the agent then incorporates on a re-run. It runs entirely on self-provisioned infrastructure: a two-server distributed deployment on Oracle Cloud with the application and database on separate hosts over a private network.
The Problem
Turning a software idea into engineered artifacts is a multi-stage process, and a single AI call cannot do it reliably or safely. A naive chatbot has no structure, no validation, no memory of prior stages, and no human oversight, so its output cannot be trusted or chained. The challenge was to build a system that decomposes the work into governed stages, enforces structured and validated output at each step, keeps a human in control of every high-impact decision, and runs long-running agent work without blocking the web layer.
Architecture & Key Work
- ▹Built a provider-agnostic model layer so every agent calls one interface using a logical model name, letting the system route to NVIDIA-hosted open models or Claude via Azure AI Foundry without changing any agent code
- ▹Designed six specialized agents, each bound to a Pydantic schema, with a self-correction loop that feeds validation errors back to the model until the output conforms to its contract
- ▹Engineered a custom orchestration engine that walks a dependency graph of stages, dispatching each stage only when its prerequisites are approved, and supporting parallel stages where dependencies allow
- ▹Implemented human-in-the-loop approval gates at every stage, including a reject-with-feedback revision loop where the agent re-runs incorporating the reviewer's natural-language changes
- ▹Ran all agent work asynchronously on Celery with Redis as the broker, so long model calls execute in isolated worker processes off the web server, with the browser polling persisted run state for live progress
- ▹Persisted all run, stage, and audit state to PostgreSQL, including an append-only audit trail that records every state transition and the identity of the human who approved or rejected each gate
- ▹Added an intent-router agent that classifies each message as a build request, a question, or off-topic, so the system converses naturally and only starts a pipeline on a confirmed build request
- ▹Built a Claude-style chat interface with conversation memory, a per-user run history, and live-rendered structured outputs including generated code files and markdown documentation
- ▹Secured the system with email-based authentication, per-user data scoping, and DRF permission enforcement on every API endpoint
- ▹Deployed the full stack as managed systemd services (Nginx, Gunicorn, Celery, Redis) across two Oracle Cloud VMs, surviving disconnects and reboots
Results & Impact
- ✓Delivered a complete, runnable multi-agent system that takes a requirement through six governed SDLC stages to validated engineering artifacts
- ✓Achieved genuine controlled autonomy: agents do the work, but no stage advances without an authenticated human approval recorded in the audit trail
- ✓Built the system to be model-agnostic, so output quality scales by pointing heavy stages at a stronger model with a one-line change and no architecture impact
- ✓Produced a detailed technical reference documenting the full request flow, from keystroke to completed run, for defensible end-to-end explanation
- ✓Runs on a self-built, encrypted, two-tier distributed deployment at effectively zero hosting cost