Hi, I'm Vinayak Mathur

Building Data Workbench, a GCP-native platform consolidating 8+ AI-powered tools for data engineering and analytics workflows across client engagements.

Architected Python/FastAPI backend services on Cloud Run with Cloud SQL and Firestore, covering user, project and job orchestration, shipped through GitHub Actions and Cloud Build with SonarQube scanning.

Designed a Pub/Sub parallel execution layer for long-running LLM workloads on Vertex AI — 100+ files processed concurrently per service with stable end-to-end latency under burst traffic.

Built a multi-agent system on Google ADK performing any-to-any SQL translation across 8 dialects, eliminating manual rewrites during client migrations.

Built a RAG service over Vertex AI vector search that auto-generates column descriptions and PII sensitivity classifications direct from BigQuery or ingested metadata.

Refactored Azure OpenAI legacy tools onto a GCP-native stack, redesigning APIs and deployment topology to fit a unified platform architecture.

Architected a production GenAI data analyst for a global beverage leader using LangGraph and Azure OpenAI, orchestrating specialised agents for SQL generation, visualisation and Pandas computation against complex business rules (YTD, market share, channel splits).

Engineered an autonomous multi-agent IT-support framework on Nvidia NIM, integrating MS Graph and a RAG debugging pipeline to triage and resolve tickets with near-zero human intervention.

Built and deployed an automated interview platform on multi-cloud Kubernetes (GCP, Azure, AWS) integrating GPT-4 and Claude — 99.9% uptime and a 40% reduction in operational cost over the prior system.

Delivered a RAG-powered audit-automation tool for a Big 4 client, cutting audit preparation time by 65%.

Trained churn-prediction models (Random Forest, GBM, neural networks), improving baseline accuracy by 15%.

Developed an anti-spoofing and speaker-verification model for on-device voice IoT using Residual Squeeze-and-Excitation networks in TensorFlow.

Achieved 4.44% EER on the ASVspoof benchmark; work published at ICCNT 2021 (IEEE).