# John D'Esposito

> John D'Esposito is the founder of Amalfi AI (https://amalfi.ai). He makes enterprise AI governed, cost-controlled and fast in production — AI GovOps, AI FinOps, AI Performance Engineering and production AI development — drawing on 20+ years of production infrastructure, API gateway and performance engineering.

- Website: https://john.desposito.us/
- Company: Amalfi AI — https://amalfi.ai/ (consulting engagements are contracted through Amalfi AI)
- Product: PromptClassify — https://promptclassify.ai/ (prompt risk classification)
- Book a 30-minute AI operations review: https://calendly.com/johndesp-amalfi
- Contact: johndesp@amalfi.ai
- LinkedIn: https://www.linkedin.com/in/amalfiai
- Location: New York City Metropolitan Area

## Services

- AI GovOps: governance as a continuous flow — risk classification, guardrails and audit evidence at the point of every AI interaction, mapped to ISO 42001, NIST AI RMF, the EU AI Act and HIPAA. Typical first engagement: a governance baseline.
- AI FinOps: every AI call — LLM, MCP, agent-to-agent — metered, attributed to an owner and reconciled, with costs reduced in-line and budgets enforced where the spend happens. Typical first engagement: an AI spend baseline from live traffic, with no application changes.
- AI Performance Engineering: latency, throughput and resilience engineered into the AI path and held to SLOs with real telemetry. Typical first engagement: a performance baseline.
- AI Development: production agents and LLM applications grounded in the client's data, with deterministic guardrails, evaluation gates and tracing from day one. Typical first engagement: a production pilot on one real workflow with a measured go/no-go.

## Featured case study: TriageBot (client: a national healthcare payments company)

A multi-agent ticket triage and diagnosis system built on the principle "the agent proposes, deterministic code decides". A six-stage LangGraph pipeline (hydrate, classify, validate, diagnose, recommend, act) kicks back incomplete tickets, classifies the rest, diagnoses root cause from point-in-time Snowflake warehouse state, and proposes remediation for human approval. It abstains when evidence is insufficient.

- Classification accuracy of 88–93% on live weekly holdouts, versus an 84% rate at which human agents agreed with each other.
- About 95% accuracy on high-confidence predictions.
- Recommendation agreement rose from 27% in early development to about 90% on production holdouts.
- No fine-tuning: it learns from human-verified precedent in retrieval.
- Regression gates over frozen evaluation sets and blind weekly holdouts; end-to-end tracing in MLflow.
- Built with LangGraph, Amazon Bedrock, OpenSearch Serverless, Snowflake, MLflow, Kubernetes (EKS), Terraform and open-weight small language models.

## Platforms

- Databricks: medallion architecture, Unity Catalog, MLflow, Mosaic AI Gateway, Lakehouse Monitoring.
- Snowflake: data-grounded agents, guarded point-in-time SQL, query-history forensics.
- Models and AI tooling: model-agnostic — frontier models (Anthropic Claude, OpenAI GPT, Google Gemini) via Amazon Bedrock and private endpoints; open-weight models such as Gemma and Qwen served locally with vLLM; every model, hosted or local, behind the same gateway governance, metering and SLOs. Builds with Claude Code, custom MCP servers and multi-model review panels.
- Gateway and observability: Kong AI Gateway, Kong Mesh and the Nginx / OpenResty engine beneath them; custom Lua plugins for authorization, prompt validation and routing; data-plane latency tuning; OpenTelemetry, Dynatrace, AWS, Terraform.

## Career

- Amalfi AI — Founder (2024–present)
- United Airlines — API platform, MLOps and AI traffic management for the mobile cloud transformation: Kong Gateway and Kong Mesh across AWS regions with custom Lua plugins, OpenTelemetry to Dynatrace, Kong AI Proxy (2021–2024)
- Nomi Health — DevOps and automation architect (2020–2021)
- Goldman Sachs — DevOps and automation architect, Apple Card cloud infrastructure (2018–2020)
- Liberty Mutual, GE Digital — DevOps and automation architect (2016–2018)
- Patents: US 8,972,569 B1 — remote and real-time network and HTTP monitoring with a real-time predictive end-user satisfaction indicator (sole inventor, 2015, https://patents.google.com/patent/US8972569B1/en); US 6,965,938 B1 — clustering servers for performance and load balancing (co-inventor, IBM, 2005, https://patents.google.com/patent/US6965938B1/en); plus a third on web-site performance monitoring and testing.
- Lafayette College — B.S., Electrical Engineering.