MCP/Agents/Skills Meetup - March 11, 2027

Join our virtual meetup to hear talks from experts on cutting-edge topics across AI, ML, and computer vision.

Connecting an agent to company data through MCP is the easy part; getting answers that are actually correct is not. Agents that query ledgers, warehouses and business systems routinely return fluent but wrong numbers: hallucinated totals, stale results and metrics computed with the wrong definition.

This talk walks through the patterns that fix it: query-first tool design instead of stuffing data into context, semantic definitions exposed as tools so agents use the same metric logic as the business, row-level citations so every figure can be traced back to the source, and eval harnesses that catch accuracy regressions before they ship. Includes code examples and a live walkthrough of an agent answering finance questions with verifiable citations.

In this short session, Adonai will show the agent built into the Voxel51 visual AI data platform working live on a fresh self-serve workspace: loading a dataset, finding and curating the right samples, labeling, and tracking down model failure modes from plain-language requests. You'll see where the agent saves the most time, where a human should stay in the loop, and how to try it free on your own data.

AI agents that call tools through MCP and read private data fail in ways demos never show: over-permissioned tools, prompt injection, answers without sources and silent regressions. This talk walks through a practical guardrail stack for production agents.

We cover least-privilege tool and MCP server design, retrieval with citations, an evaluation suite that runs in CI, human approval gates for risky actions, and observability for audit. Attendees leave with a checklist and reference architecture they can apply to their own agents.