shivadeepak shanigaram

I build LLM systems that hold up in production by enforcing correctness in code—not prompts—and backing decisions with real evaluation. My focus is shipping agents that behave reliably under messy, adversarial, real-world inputs.

As an AI Engineering Intern, I built and shipped a production natural-language search agent for plain-English questions over a company graph database. I implemented LangGraph + Gemini tool-calling with streaming responses and deployed it on AWS.

I also designed a deterministic security layer for safe multi-tenant database querying, making it structurally impossible for one customer’s queries to access another customer’s data. To improve agent quality, I built internal benchmarks and raised correctness from 64% to 84% through model comparisons, prompt iteration, and hand-audited failure analysis.

Beyond agent behavior, I strengthened the data and cost envelope: I rebuilt the ETL pipeline to eliminate silent data loss (100% field capture) and lifted date extraction accuracy from 75% to 96%. I reduced model spend and improved reliability using a rules-based filter (no LLM call for trivial/malicious inputs) plus rate limiting, per-customer quotas, and timeout/fallback handling.