Nikunj Khitha

Integrated National Power Portal data into a national renewable energy dashboard serving 150+ power stations and improved reporting accuracy by 30%.

Built a secure Java/PostgreSQL file management system with role-based access control that improved retrieval efficiency by 25% across 5,000+ files.

Developed a MERN conference room booking system that cut booking time by 60% and reduced scheduling errors by 40%.

Built Node.js and FastAPI microservices, improving API performance by 40% and reducing deployment time by 42% with Docker and GitHub Actions.

Engineered Fantasy GPT with RAG, LangGraph, backend APIs, agents, and DeepEval quality checks to resolve 98% of complex sports queries.

Built Python-based ETL pipelines to collect sports data from multiple sources and ingest it into MS SQL for Fantasy GPT SQL RAG workflows.

Worked across AIKO, a voice-based sports companion using Azure Speech SDK for text-to-speech and speech-to-text, user-level personalization, and live AI-generated commentary in 20+ languages.

Built AIKO personalization features for on-the-fly highlight reels, where AI agents stitched sports moments based on each user's profile and interests for a product presented at IBC 2024 in Amsterdam.

Created the initial Java framework for Anya, ArmorCode's platform agent; designed, implemented, and owned its short- and long-term memory layers; and set up Langfuse-backed evaluation for agent accuracy and memory behavior.

Ported 2 of 6 Anya sub-agents from LangChain4j to Spring AI, reimplementing memory and evaluation integrations for the new architecture.

Owned backend integrations for 5+ security tools, including Black Duck, Snyk, and Checkmarx, on an AppSec platform aggregating findings across 130+ connectors.

Created AI-assisted code generation utilities with template engines and AST parsing to automate new integration scaffolding, reducing per-integration boilerplate setup time by 30%.

Co-built Agentic Office OS, the internal platform of autonomous, human-triggered, and human-in-the-loop agents reachable org-wide through Slack, translating ambiguous requirements from an external business consultant, executives, and internal teams into agents, skills, MCP integrations, and approval workflows for 20+ business functions, then owning rollout, debugging, and enablement.

Built an OpenTelemetry-based LLM evaluation platform on Langfuse so every prompt, model, agent, skill, and MCP tool change ships on measured accuracy, latency, and cost. It scores 9 AI surfaces against golden datasets using deterministic checks, scikit-learn classification metrics, Ragas RAG scores, and LLM-as-a-judge graders validated at Cohen's kappa >= 0.7, enforced as Jenkins CI gates.

Delivered 10+ production MCP servers in ArmorCode's shared enterprise tool registry under multi-tenant AppSec constraints, implementing OAuth2/RBAC controls, tool-level permission tiers, explicit denial behavior, and audit attribution, with 20 of 20 authorization checks validated across three access tiers.

Govern company-wide model and MCP access as sole maintainer of a LiteLLM gateway, issuing scoped API keys with per-model spend budgets and distributing RBAC-gated MCP servers to employees' Claude Desktop via an .mcpb proxy. Cut recurring LLM spend by tracing 57% of gateway cost across 50,000+ requests to 10+ automations, then migrating models and splitting system/user prompts to enable Bedrock prompt caching, while flagging a deterministic rewrite worth a further ~95% reduction.

Built ArmorCode's tenant-scoped knowledge-graph RAG layer over 1M+ entities of root-cause analyses, test cases, and product documentation in Neo4j and pgvector, giving Office OS agents grounded product knowledge without cross-tenant leakage, with retrieval scored on accuracy, ranking quality, and context precision across five query modes.

Cut CS and support escalations to engineering with a codebase-search MCP service in Go over 8 product repositories, fronting a read-only agent against a daily-reindexed vector index. Restored it after 100+ queries returned zero content by isolating two independent causes in SDK and host internals: a Go output-schema defect that made schema-aware clients discard every answer, and a query deadline sized for a longer client timeout tier. Shipped the fix with a negative-control test and flagged 14 exposed connectors.

Built AI documentation automation that triggers when a Jira ticket is completed: an agent creates or updates documentation in GitHub and opens a pull request, reviewers use Quill (Electron/React/TypeScript) to review and edit the changes, and merged documentation syncs to Zendesk. Built Quill as the workflow's review app with a WYSIWYG editor, an embedded agent terminal, per-branch Git worktrees, and permission-aware GitHub actions.