Building agentic systems in production — open to new work
Applied AI Engineer. I take multi-agent systems from prototype to production inside large, regulated organizations — and make their behavior legible enough that leadership will actually turn them on.
The interesting part of applied AI isn't getting a model to answer — it's deciding what the system does when it isn't sure. This is the shape most of my production work takes.
Scroll the diagram sideways on a small screen
LangGraph and LangChain multi-agent graphs with parallel retrieval, confidence-based routing, and human-in-the-loop checkpoints placed exactly where being wrong is expensive. Function calling against real systems of record, with code-enforced safety gates around anything that writes.
Async FastAPI inference endpoints with schema-validated structured outputs, prompt and context engineering that survives real inputs, caching and concurrency control, and per-task cost and latency budgets that hold up when volume arrives.
Hybrid dense + BM25 search with cross-encoder re-ranking and semantic chunking, tuned against a golden set rather than vibes. Citation-grounded generation, multi-tenant isolation, and dead-letter-queue recovery so ingestion failures are reprocessable instead of silent.
Golden datasets, LLM-as-judge scoring, faithfulness checks, and regression evals that catch retrieval and prompt drift before release. Full agent tracing plus per-agent KPIs — cycle time, resolution rate, cost per task — so agent behavior can be explained to people who don't read traces.
Agentforce and Salesforce-connected automation where the CRM stays the system of record and the agent does the work around it — bi-directional integration across Sales, Service, Experience and Data Cloud, with exceptions routed back to a human queue instead of guessed at.
Spark and Azure Data Factory pipelines feeding warehouse, reporting and retrieval layers, plus natural-language querying over large operational datasets with dynamic schema injection and static SQL validation — read paths open, destructive paths closed.
Containerized agent services on Azure with autoscale, gateway policy, managed identity, secrets handling and queue-based decoupling — shipped through CI/CD with telemetry wired in from the first deploy rather than bolted on after the incident.
Client and employer names withheld. Detail available on request.
End-to-end owner of an Azure multi-agent system validating legacy circuit records across seven OSS/BSS systems to de-risk network decommissioning. Parallel retrieval, confidence-based routing, entity resolution across inconsistent identifiers, production-safe NL2SQL, and domain compliance guardrails that keep autonomous decisions inside regulatory boundaries.
Validation cycle time: hours → minutesA Salesforce-connected agentic platform handling multilingual purchase-order extraction, SKU matching, compliance validation and automated order creation — replacing a manual multi-team process with an agent-driven one, with human-in-the-loop exception handling and hardened ingestion for regulated data.
Manual multi-team process → automated workflowGolden datasets with LLM-as-judge scoring plus end-to-end tracing, paired with hybrid dense + BM25 retrieval and cross-encoder re-ranking. Retrieval and prompt regressions started getting caught before release instead of in production.
Higher answer relevance · faster failure diagnosisArchitected and delivered automated submission, tracking, eligibility review and approval workflows under federal compliance constraints, on reusable platform patterns several programs could share. Backed by Spark and Azure Data Factory pipelines into the analytics layer.
Manual case-intake time down 60%Eight-plus years across the lead-to-revenue lifecycle — CPQ quote-to-cash, contract lifecycle management, order management, marketing automation and customer service — which is why the AI work lands on systems that already exist rather than beside them.
Quote turnaround +30% · contract generation −45% · adoption +40%Agentic AI, retrieval, evaluation, or the systems integration underneath all three.