Applied AI systems for engineering teams.

The challenge

A load-balancer API exposes hundreds of object types and validation rules. The work of translating intent into a valid request was getting in the way of the actual engineering decision.

What I did

I created an open-source MCP server that uses the Avi load balancer’s Swagger schema to turn natural-language requests into appropriate API calls. When a call fails, it uses schema context to help the model recover rather than returning an opaque error.

The challenge

Customer support bundles contain logs, configuration, and signals spread across many files. Investigating them manually was slow and difficult to repeat consistently.

What I did

I designed an agentic system that reads a support bundle the way an experienced engineer would: it indexes evidence in Qdrant, reasons across logs and configuration, and produces a plain-English root-cause report. I also added observability, PII scrubbing, and group-based access control so it could operate in an enterprise setting.

The challenge

Text similarity alone cannot explain how a distant function contributes to an issue. Engineers needed a system that could follow code structure, not just matching words.

What I did

I built a code-intelligence layer using tree-sitter parsing across Go and Python. It indexes more than 30,000 symbols and combines exact matches, LLM query expansion, semantic search, and call-graph traversal.

The challenge

Operating a load balancer well takes real expertise — reading application analytics, diagnosing performance issues, and knowing the right sequence of steps for upgrades and day-to-day configuration. Most of that knowledge lived in documentation and in people's heads, not in the product itself.

What I did

I built a conversational GenAI assistant for Avi that turns its built-in application analytics — including the application health score — into plain-English explanations, and answers operational questions directly from Avi's public documentation. Administrators get instant KB-backed answers, natural-language summaries of performance issues, and guided recommendations for tasks like upgrades, instead of working through manual steps by hand.

The challenge

Product users had separate places to look for documentation answers, live system state, and training material—none connected the full picture.

What I did

I built a multi-source RAG assistant that combines documentation Q&A with live controller introspection through function calling and relevant training-video recommendations.

The challenge

Jenkins failures were being sorted one at a time, even though many belonged to known patterns and existing issues.

What I did

I built a machine-learning pipeline that predicts failure categories from log patterns and auto-assigns confident matches to the related Jira issue. It also supports crash reporting and hybrid cloud/on-prem data sync.

The challenge

A failed automation test comes with a wall of logs and a large test-automation codebase behind it. Working out whether a failure is a real product bug or a flaky test still meant an engineer reading through both by hand.

What I did

I'm building an agentic system that reasons over the exact test-run logs and the automation code together, rather than guessing at file and line numbers. It's given tools to inspect the real source (resolving symbols and dependencies as the code actually defines them) and grep prior logs, then a separate critic agent checks the explanation is grounded in that evidence before it's shown to anyone. It's deployed as a queue-based service that plugs into our existing CI pipelines.