Research & innovation
Research direction and ongoing practice. No academic publication on this topic yet.Autonomous AI Agents for End-to-End Software Development
The questions I'm pursuing — and where I already practise them.
How far can a team of AI agents take a product — from a vague requirement to a tested, deployed system — and what has to stay deterministic for that to be safe? My interest is in the engineering around the model: planning, verification, permissions and evidence.
01Research agenda
- 01
Multi-agent collaboration
Splitting software work across specialised agents — requirements, architecture, implementation, review — and the protocols that let them hand work to each other.
Open questions
- When does a graph of agents outperform a single strong model with tools?
- What should an agent card or contract contain to make delegation verifiable?
In practice
- 02
Code generation with validation
Generated code is a proposal. Tests, static analysis, benchmarks and screenshot diffs are the evidence that turns it into a change.
- 03
Permissions, audit and responsible autonomy
An agent that can act needs a deterministic gate: risk levels, on-screen confirmation, redaction and a tamper-evident trace of what it did.
- 04
Self-healing systems
Systems that detect their own failures and recover — from replayable webhook pipelines to models that are retrained when the attack they face is new.
Open questions
- What is the minimum evidence an agent needs before applying its own fix in production?
In practice
- 05
Requirements analysis with AI
Turning a founder's brief into phases, backlogs and decision records that humans can review — the part of software that fails most often.
Open questions
- Can an agent tell an ambiguous requirement from a missing one?
In practice
02Papers & reports
- DoS Detection & Response
Course research paper — Applied Machine Learning for Security Analytics · 2026
ML-Based DoS Attack Detection and Automated Response (CIC-IDS2017)
Random Forest detector with F1 0.9989 on held-out traffic; evaluates generalisation to an unseen DDoS (recall 0.633) and shows a strike-based response layer that still stopped 92.3% of attack flows in simulation.