AI Automation

Agentic workflows that do the work — not just chat about it.

The wins from GenAI come from automating end-to-end workflows, not from another chatbot. I design agentic systems that combine an Enterprise Knowledge Graph, MCP tool catalog and on-prem LLMs to take real action — with humans in the loop where it matters.

Workflows I automate

  • Document review & extraction — contracts, RFPs, regulations, policy docs → structured fields, comparisons and red-flag reports.
  • Complaints & case triage — classify, route, summarise and draft responses; escalate edge cases to humans.
  • Inspection & field operations — generate inspection plans, parse photos and notes, populate work orders.
  • BI & analytics copilots — natural-language Q&A over Power BI, SQL warehouses and KPIs with row-level security.
  • Knowledge worker assistants — research, briefing, drafting and meeting prep grounded in your knowledge graph.
  • IT & ops automation — ticket triage, runbook execution, change-request drafting via MCP tools.

Architecture pattern

Planner LLM → MCP tool calls (KG search, ERP, BI, ticketing) → Critic / verifier → Human-in-the-loop checkpoint → Action. Built with LangGraph or Semantic Kernel, evaluated with golden datasets, instrumented with OpenTelemetry, and deployed on your own GPUs.

Guardrails & governance

  • Deterministic tool schemas — no free-form SQL or shell.
  • Policy filters and PII redaction at retrieval and output time.
  • Per-user, per-tool authorisation mirrored from source systems.
  • Full audit log: prompt, tools called, inputs, outputs, decision.
  • Eval harness gates every prompt / model / tool change.

Outcomes I aim for

  • 30–50% reduction in manual handling time on the target workflow.
  • Measurable accuracy & latency SLOs, not vibes.
  • Reusable MCP capabilities the rest of the org can build on.