Architecture I’ve led — the problem, the decision that mattered, and what I was responsible for. Two in full, because depth is the point and a shallow case study proves nothing; the rest in brief.
Most of my current work runs through Syntphony, NTT DATA’s productized technology ecosystem — specifically Syntphony AI, Syntphony Open Telecom Networks, Syntphony Insurance Cloud Migration, Syntphony Stations, and Syntphony Commerce & Payments.
A national insurance data ecosystem aggregates sensitive records from every insurance company in the country. Perimeter security doesn’t address insider threat or an AI-era access pattern: autonomous agents, not …
Mainframe modernization normally starts with senior COBOL engineers spending weeks manually reading code to reconstruct business logic and map dependencies — slow, expensive, and inconsistent, and that expertise is …
A reusable multi-agent platform layer instead of another per-engagement build: MCP as the tool-integration primitive, A2A for cross-domain agent requests. Domain agents plug in and inherit the governance automatically — per-agent OAuth2 identity, full tool-call audit logging, per-agent and per-tool rate limits, a human-in-the-loop approval gateway for high-risk actions, and OpenTelemetry tracing per agent turn. The governance layer is the actual deliverable; the reuse follows from it.
Two distinct paths rather than one pipeline, because the query types differ. Live metrics are answered by querying semantic models directly with row-level security passed through as the calling user — never embedded, because a stale number is worse than no answer. Document queries run through Azure AI Search over permission-scoped content. An LLM classifier routes each question, or both ways when a query spans metrics and documentation.
Five OSS domains — RAN, Core, Transport, Inventory, Assurance — each with its own data model, API and metric vocabulary, which makes single-agent query planning unreliable. A query-planning agent decomposes the question and dispatches to domain-specialist agents, then a correlation layer assembles an answer cited by system, network element, metric and timestamp rather than letting the model summarise ungrounded.
Conversational analytics embedded in the client's existing back-office system, with Semantic Kernel routing intent to specialist agents over existing Power BI semantic models rather than a rebuilt analytics layer. Multi-tenancy through Entra ID plus row-level security scoped per business unit and country, so one assistant serves every entity without duplicating access-control logic.