Manidheer Reddy Nalla

• Built a natural language search interface over the state's biological survey database (SQL Server, 1M+ records) using Azure OpenAI and RAG, so scientists can ask plain-English questions instead of stacking grid filters • Own the retrieval layer end to end: chunking strategy for survey reports and species metadata, embedding generation, Azure AI Search indexing with program and county filters, and a nightly sync job that picks up new SQL Server records • Wrote the Python FastAPI service sitting between the LLM and the existing .NET 8 Blazor app, handling streaming responses over SignalR, retries, token budgets, and Entra ID auth end to end • Curated an eval set of about 250 Q&A pairs with agency biologists and wired it into the release process; it caught two retrieval regressions and settled a chunking debate with numbers instead of opinions • Cut monthly Azure OpenAI spend by roughly 40% through response caching and routing simple lookups to GPT-4o mini, reserving the larger model for multi-record synthesis • Added guardrails required for government use: PII redaction on inputs, prompt injection filtering, and citations grounded to source records so answers can be verified before they land in official reports • Develop core features of the Blazor Server application itself: DevExpress data grids with server-side filtering and paging over large survey datasets, reusable Razor components, and role- and program-based field visibility tied to Entra ID groups • Extend the internal ASP.NET Core REST API and data layer (EF Core, Dapper, LINQ, stored procedures) against SQL Server, and contribute code reviews and unit tests as part of a small development team

• Fine-tuned a DistilBERT classifier (PyTorch, Hugging Face) to route incoming legal documents across 60+ practice areas at 0.89 F1, replacing a keyword rules engine the team had been patching for years • Built NER pipelines (spaCy plus custom extraction heads) pulling parties, citations, courts, and filing dates out of case documents to enrich the search index • Maintained the batch infrastructure around the models: Airflow DAGs, PySpark preprocessing over multi-TB document dumps, and experiment tracking in MLflow so training runs were reproducible • Served models from Docker containers behind a REST API on AWS EKS, holding p95 latency under 200ms during peak document ingestion • Developed asynchronous FastAPI endpoints with Pydantic validation to improve document ingestion throughput and reduce API latency under Kubernetes orchestration. • Implemented automated retraining pipelines using Celery and Python to maintain model accuracy on evolving legal business data sets. • Integrated conversation status tracking within AI co-pilots to enhance user interaction workflows and streamline case document processing.