Ade
Daramola

GitOps Sentinel is an autonomous remediation platform for Kubernetes clusters. When Prometheus fires an alert, Sentinel doesn't just send a PagerDuty notification — it intercepts the anomaly, reasons through root cause using a multi-agent pipeline, and resolves the incident by committing a fix to Git. Argo CD picks it up, reviews it against policy, and applies it. The cluster heals itself through the same controlled write path a human engineer would use.

The design comes from a real operational insight: not all incidents are equal. A pod OOMKilled on a non-critical service at 3am doesn't need a human. A cascading failure across stateful workloads does. Sentinel gates every remediation on a confidence score — routine incidents resolve automatically in seconds, while ambiguous or high-risk scenarios escalate to on-call before anything touches the cluster. If a remediation fails, it triggers an automatic revert. The cluster is never worse off than it was before Sentinel ran.

This is the kind of system that pays for itself the first time it silently fixes something at 4am that would have been a two-hour incident.

MedQuery is a medical Q&A system built around the assumption that in a high-stakes domain, a wrong answer is worse than no answer. Before any query reaches a retrieval step, a safety guardrail classifies it — rejecting non-medical and high-risk queries before any LLM processing runs. This isn't a content filter bolted on at the end; it's the first thing that executes.

When a query passes the guardrail, it's routed across three source tiers — a structured PubMed/FDA medical corpus, FDA drug data, and live web search — with up to three relevance checks before a response is generated. The system knows whether it found enough grounding evidence before committing to an answer. Answers stream token-by-token with source-quality labels attached, so a clinician or patient always knows whether a response is backed by peer-reviewed data, an FDA record, or a web result. That distinction matters enormously in practice.

The infrastructure runs fully on AWS — ECS Fargate, RDS with pgvector, CloudFront — deployed via Terraform. No manual steps, no snowflake servers.

The Multi-LLM Platform is an inference gateway that puts OpenAI, Anthropic, and AWS Bedrock behind a single unified API. Application code calls one endpoint and never knows — or needs to know — which provider handled the request. The gateway routes based on request complexity, cost threshold, and caller latency SLA. Simple requests go to cheaper models. Tight-SLA requests go to faster tiers. The routing logic is policy-driven, not hardcoded.

Provider health is monitored in real time. When a provider degrades, it's removed from the active pool immediately — not after a timeout, not after a human notices. A semantic cache layer sits in front of the routing logic and returns results for prompts that are similar enough to previous requests, which means a meaningful percentage of production traffic never hits an LLM at all. That translates directly to lower cost and lower latency.

Every request emits structured telemetry — tokens consumed, model selected, latency, cache hit rate, and cost by provider. If you're running an AI product without this level of visibility, you're operating blind. This platform is what it looks like to operate LLM infrastructure the same way you'd operate any other production service: with observability, redundancy, and cost accountability built in from day one.