Osama Malik
Home / Blog

This is the blog of Osama Malik (also written Usama Malik), an AI agent engineer at Onlinx Solutions. The posts are first hand notes on BleedStop, a merged Pipecat fix, and the prompt injection layers in BleedStop.

Blog

Notes from the work

Three posts. Each one uses the public repo, the case study, or the pull request itself. I am not adding numbers that those sources do not already publish.

BleedStop: How I Built an LLM Cost and Security Guard

BleedStop is a control plane I built so an app asks before it spends. It answers ALLOW, LIMIT, CHALLENGE or DENY, with Redis budgets and a Postgres audit log. The measured numbers, and the ones I will not stretch, are in the post.

Pipecat PR 5760: Function Call Timeouts Kept Armed

In Pipecat, an intermediate function call result was cancelling the tool timeout. The merged fix leaves the deadline armed until the result is final. This is the story of that pull request.

Prompt Injection Defence Layers I Use in BleedStop

BleedStop puts a phrase list in front of an optional DeBERTa classifier, then a budget decision. The held out classifier caught 77.5% at a 5% false positive rate. The phrase list on that same set is much weaker, and I say so.

Related pages: home, BleedStop case study, Agent Receptionist, open source, hire.