Vikas
Rimmalapudi

Challenge

Manual incident response was slow, repetitive, and error-prone during critical outages across 200+ servers.

Solution

Designed multi-agent system with LLM-guided runbook execution, confidence-based escalation, and safety guardrails to prevent unsafe automated actions. Implemented autonomous diagnosis agents that analyze logs, metrics, and system state to identify root causes.

Results

• Cut mean time to resolution (MTTR) from 45 minutes to 8 minutes
• Reduced alert noise by 90% through intelligent filtering
• Automated 80% of recurring remediation tasks
• Zero unsafe automated actions in production

Tech Stack

Challenge

Users needed fast, accurate answers from massive, unstructured knowledge bases spanning 50K+ documents.

Solution

Built hybrid retrieval architecture combining dense and lexical search with reranking. Engineered high-throughput ingestion pipeline processing files up to 1GB in under 60 seconds using parallel processing, adaptive chunking, and batched embeddings. Migrated from serverless to self-managed GKE for cost efficiency.

Results

• Serving 10K+ daily active users
• Sub-3-second p95 query latency across 50K+ documents
• Reduced hallucination through hybrid retrieval
• Cut annual infrastructure costs by $40K

Tech Stack

Challenge

Sales teams needed immediate, actionable coaching feedback from customer calls at massive scale.

Solution

Led team of 4 engineers to build distributed real-time GenAI system. Created LLM coaching agents evaluating transcripts against configurable rubrics. Implemented streaming and parallel inference pipeline with backpressure handling for stable latency under high-throughput workloads.

Results

• Processing 130K+ calls per day
• Sub-second end-to-end latency
• Delivering structured feedback to 500+ sales reps
• Comprehensive observability with LangSmith and Langfuse

Tech Stack

Challenge

Manual case reviews created bottlenecks and inconsistent quality standards for enterprise support operations.

Solution

Built AI-powered evaluation platform on Azure Functions and Vertex AI. Engineered event-driven pipelines using Azure Service Bus with parallel LLM inference, independent timeout groups, and retry logic ensuring zero downstream reviewer blockage at scale.

Results

• Automating quality review of 500-2000 cases daily
• Multi-cloud architecture (GCP + Azure) provisioned via Terraform
• Zero downstream blocking through intelligent retry logic
• Consistent quality standards across all reviews

Tech Stack