Microsoft Xbox · Tech Mahindra · CodeRush 2026 Finalist
An AI agent for handling production problems. When an alert comes in, it sorts it, looks up similar past incidents, suggests the likely cause, and drafts a fix, but a person reviews and approves every change.
POC complete, intelligence layer in progress. My own concept and architecture; full implementation details shared on request.
Problem
When something broke, engineers created tickets by hand hours later, around 30% of incidents went to the wrong team, and real investigation only started after 20 to 40 minutes of reading logs. The same problems were investigated again and again, because nothing remembered past fixes.
Design Decisions
Grounded in real history. The agent only suggests a fix if it can point to a similar past incident or a runbook (a step-by-step guide for a known problem). This prevents the most common AI mistake: confidently inventing a wrong answer.
Knows when to step back. If it is not confident enough, it hands the incident to a person instead of guessing.
A person always approves. Nothing is merged automatically; an engineer reviews and owns every code change.
Protects sensitive data. It removes secrets and private data before anything is sent to the AI model, and logs every action.
Controls cost. Simple sorting uses a small, cheap model; deeper reasoning uses a bigger one, with a spending cap per incident.
How It Works
One pipeline takes a raw alert all the way to a draft fix, then feeds the result of every closed incident back in so it keeps getting better.
It takes in alerts from monitoring tools and reports from users.
The AI reviews similar past incidents and guides, found using both keyword and meaning-based search.
It sets the severity, picks the right team, and suggests a likely cause, along with a confidence score.
It opens the ticket, creates a code branch, and drafts the fix. It never merges on its own.
The outcome of every closed incident is saved so future suggestions get better.
Engineering Challenges
An AI that touches live incident handling has to satisfy a reliability and security reviewer before anyone will trust it.
A person must review and approve every change before it goes in; nothing applies automatically.
It can only suggest something it can back up with a real source, plus a confidence score that sends shaky cases to a person.
Secrets and private data are stripped out first, and the AI models do not train on the input.
Cheap tasks use a small model and only deeper work uses a bigger one, with a spending cap per incident.
Built With
An AI agent layer over Azure OpenAI and smart search, with ongoing quality checks and tracing for every recommendation.
Impact
<10 min
Alert to draft fix (proof of concept)
95%
Routing accuracy (proof of concept)
Every change
Human review
What This Shows
The interesting part is not calling an AI model. It is the grounding in real history, the confidence checks, and the rule that a person approves everything, which together make the automation trustworthy. The numbers here are from a proof-of-concept, not a full company-wide rollout.