Sai Anuhya Bandi

About

Backend/data engineer who built an enterprise LLM assistant (AI Genie) at Broadband Insights using a LangChain + GPT-4 + Pinecone RAG pipeline to automate broadband analytics reporting. Developed Python/Dagster ETL processing 10M+ records/day and improved data freshness by 60%, with production-grade scalability patterns (async workers, containerized microservices, Kubernetes) and strong multi-tenant isolation practices.

Experience

Education

Key Strengths

• Built enterprise LLM assistant (AI Genie) using LangChain + GPT-4 + Pinecone RAG for broadband analytics insights
• Designed and operated ETL pipelines processing 10M+ records daily
• Improved data freshness by 60% through pipeline and processing optimizations
• Backend scalability patterns: modular Flask architecture, async workers to prevent API blocking
• Production deployment scalability via containerization and Kubernetes (horizontal scaling, rolling updates, minimal downtime)
• Database performance optimization with SQLAlchemy/PostgreSQL (indexing, connection pooling, bulk ops, avoiding N+1 queries)
• Multi-tenant design with strong data isolation (schemas/tenant IDs), access controls, encryption, and quotas/rate limits
• Integrated AI/ML workflows using microservices and event-driven processing across cloud-native stack (Kubernetes/Lambda)
• Built and operated production Python backend services for LLM/workflow-driven analytics (FastAPI/Flask) on AWS/GCP with Docker/Kubernetes
• Debugged and resolved P95 latency spikes/timeouts by addressing PostgreSQL connection/write pressure, indexing, batching writes, and adding bounded async concurrency/backpressure
• Strong production-grade AI/LLM guardrails: RAG from approved sources, structured JSON + schema validation, and post-generation verification against databases to prevent hallucinations
• Designed reliable multi-step agent systems with clear planning vs deterministic execution, tool timeouts, idempotent retries, and stop/escalation logic
• Improved maintainability and reliability via typing (MyPy), Pydantic schemas, modular boundaries, and tests (unit/contract/integration); fixed async race condition by removing shared mutable state
• Relational data modeling for workflow systems (runs/steps/artifacts) with constraints for idempotency and performance-focused indexing
• Built and deployed production LLM/RAG analytics platform for executive-level broadband analytics
• Scaled data processing via ETL pipeline handling 10M+ records daily
• Improved LLM reliability at scale via grounding (schema-aware prompts, metadata filtering) and retrieval tuning (chunking)
• Reduced latency through caching and parallel retrieval
• Strong evaluation/monitoring approach using golden datasets and production metrics (correctness, latency, retrieval precision, fallback rates)
• Effective collaboration with non-technical stakeholders through prototyping, validation, and iterative delivery