Agentic AI Developer & LLM-Powered Systems Engineer

Most business questions that require querying a database, cross-referencing multiple systems, or synthesising an answer from live data still route through a person: an analyst who writes the SQL, a developer who builds a one-off script, or a manager who waits for a report. That works until the question is ad hoc, the data is large, or the person isn't available. The bottleneck isn't the data. It's the manual translation layer between a plain-English question and an executable action.

I build agentic systems in LangGraph, modelling the agent's reasoning as an explicit state graph rather than a single prompt-and-response loop: the agent plans a step, selects a tool, executes it, observes the result, and decides whether to continue or respond. This makes the reasoning inspectable and debuggable instead of a black box. For Text-to-SQL specifically, that means schema-aware prompt context, query validation before execution, and hard guardrails: read-only enforcement, row limits, and retry caps on generation failures, so the agent can't run an unsafe or runaway query. Where a task needs both internal data and external context, I combine structured database retrieval with live web search in the same reasoning loop. For anything conversational or long-running, I add a persistent memory layer that stores entities, events, and prior context across sessions rather than treating each request as stateless. I also have working knowledge of the Model Context Protocol (MCP), tool schema design and context boundaries, which shapes how I structure tool interfaces when connecting an agent to external systems.

A typical delivery is a LangGraph agent deployed behind a FastAPI service, streaming the agent's reasoning steps over Server-Sent Events as they happen, so users see what the agent is doing (which tool it's calling, what it's querying) rather than staring at a blank loading state, followed by the complete synthesised report once the agent finishes. Tool-calling is wired directly to your database or internal APIs, with safety guardrails enforced at the query layer. This is the same architecture behind my AI Sales Intelligence System: a ReAct agent that turns plain-English questions into validated, read-only SQL over a 50M+ record PostgreSQL sales warehouse, blends the results with live web search for competitor benchmarking, and streams its reasoning live before delivering the final synthesised answer, with no analyst in the loop.