I build production LLM agents and the automation around them: eval-gated support and CRM agents, an internal MCP layer over the business systems, call and content QC pipelines. Most of my public work is security and tooling for the agent ecosystem, since that is the part everyone is shipping fast and checking last.

Now, at i-sdd (Be Z Online): support, marketing, and sales-ops automations on Azure AI Foundry. Took one multilingual support agent's answer-quality pass rate from 24% to 64% under an eval gate, and built the internal MCP server the agents and analysts run through.

Public projects

• mcp-guard — adversarial fuzzer for MCP servers. 65 findings against Anthropic's reference servers, published as a leaderboard.
• sqltok — schema token-budget manager for Text-to-SQL. Keeps 97% of gold tables at 2,000 tokens on BIRD.
• agenteval-bench — pytest-style eval suites for agents, with CI gating on regressions.
• protobuf-fuzz-guard — CVE-pattern scanning and fuzz harnesses for .proto services (Python, C++, Go).
• altius — PE risk and forecasting (de-smoothing, Chronos, AutoGluon). Live dashboard.

Stack: Python, TypeScript, SQL · Azure (Foundry, Functions), AWS · MCP, A2A, RAG, LangGraph · PyTorch, LightGBM, QLoRA · Docker, Kubernetes, Bicep.

A lot of the work is agentic security across the whole stack: code, DevOps pipeline, deployed infra, PII in transcripts, and agent-token guardrails. Open to research collabs and MSc discussions.

Reach me: LinkedIn · shovalb9@gmail.com