I architect enterprise AI systems — and own them end to end.

[problem]

One retrieval strategy couldn't reliably serve the range of enterprise questions. Compliance, product, facility, network, and pricing queries each needed a different approach — without sacrificing latency, resilience to partial failures, or a safe rollout path.

[architecture]

An LLM router classifies intent into one of five domains through an explicit state machine. Every routing decision lands in an enumerated, testable outcome. Domain agents retrieve in parallel, results merge through Reciprocal Rank Fusion, and timeouts preserve partial results instead of zeroing out the response.

Chose deterministic routing over free-form agent planning — traded flexibility for auditability and testability.

One tradeoff was accepted deliberately: a measured +7% in end-to-end latency bought an 18% reduction in response-generation time. That's the part of the wait users actually feel.

[measured]

• 88% measured routing accuracy [measured]
• Fallback to the legacy path reduced from 32% to 0% after a router fix [measured]
• Parallel router classification saves 500–1,500 ms per query [measured]
• 360 automated tests [verified]

[problem]

RAG pipelines tend to retrieve too much, too loosely, wasting context budget and raising hallucination risk. Enterprise content needs exact matches and fuzzy ones, plus relationships that vector similarity alone can't see.

[architecture]

Hybrid retrieval combines vector and keyword search, behind layered query understanding: rewrite, decomposition, entity extraction, and geographic entity resolution. A GraphRAG enrichment layer traverses knowledge-graph relationships with typed seed extraction and relevance scoring, and character-budget management compresses what finally reaches the model.

Retrieval and reasoning are separate layers — each gets its own controls, ranking, and evaluation.

[measured]

• Graph-guided filtering narrowed 80 candidate sites to 3 relevant matches — a 71% reduction in context size [measured]
• Geographic resolution maps 68 cities, validated by a 43-case stress bench with 99 assertions [verified]

[problem]

Leadership-driven platform consolidation required integrating a production Python AI backend into a broader monorepo — with consistent LLM client behavior, robust persistence, and no silent regressions along the way.

[architecture]

A FastAPI backend with async persistence on PostgreSQL and pgvector, LLM orchestration unified through Semantic Kernel. Structured outputs enforce schemas at every trust boundary, more than eight feature flags gate each new capability, and shadow-comparison tooling runs the legacy and new paths side by side with per-prompt deltas and aggregate reports.

Testing is architecture — the test suite is the contract that lets the platform change safely.

[measured]

• 2,260 automated tests passing (12 skipped) in this integration alone [verified]
• Part of a 3,000+ test surface across the platform ecosystem [verified]

[problem]

The platform needed an evidence-based path for a potential port to a broader internal platform, spanning authentication, storage, retrieval infrastructure, and feature parity. None of it could rest on assumptions.

[architecture]

A verification-first assessment, run in phases: discovery, foundation, core capabilities, feature and UX, operations, synthesis. Every architectural decision is logged with an audit trail: authentication strategy, event-driven text extraction, provider selection. Stakeholders can follow the reasoning that produced each one.

Architecture is a communication job — decisions are logged, presented, and defensible.

[measured]

• Architecture decisions logged with an audit trail; phased assessment in progress [documented]

[problem]

Every team that wants to build with the organization's AI models needs credentials and quotas, and somebody has to be accountable for both. Ad-hoc key sharing doesn't survive an enterprise. Developers needed self-service access; the platform needed governance that can't be bypassed.

[architecture]

My own creation, designed and built from the ground up as lead: a governed AI gateway that brokers the organization's models behind one managed upstream credential consumers never see. Developers request access through a portal; approval is the sole gate; provisioning flows through a transactional outbox with a reconciler that detects drift and cancels anything that tries to route around the approval path. Keys are generated once, revealed once, rotated on demand — and every subscription is metered, with token usage rolled up into per-team cost showback.

One managed credential upstream, zero shared secrets downstream — access is a governed grant, not a copied key.

[delivered]

• Request-to-provision flow verified in production [deployed]
• Least-privilege by construction: provisioning roles can create access but can never read keys [confirmed]
• Per-subscription usage metering and cost showback, priced from a versioned rate book [deployed]

[problem]

An AI agent in beta is a demo with potential. Shipping it inside Microsoft meant showing how it behaved across the range of things real users would ask, before anyone dropped the flag.

[built]

My largest contribution was the agent's evaluation platform, originated and built end to end: independent quality instruments whose deltas had to agree before a result counted, a harness that exercised the real published agent through the full production path rather than the API, and the statistical discipline — paired tests, false-discovery control, practical-significance thresholds — to tell genuine regressions from noise.

A beta becomes a product when you can prove how it behaves.

That evidence became the release gate: I authored and published the go/no-go readiness assessment behind the production rollout.

[delivered]

• The agent shipped: beta flag removed, streaming rolled out to production [delivered]
• Go/no-go readiness assessment authored and published — verdict: ready [delivered]
• Evaluation platform originated and built [confirmed]
• Direct engagement, April–July 2026 [confirmed]