Newton (Nick) Duong
AI/ML Engineer & Technical Architect
Most recently Senior Staff Software Engineer at The Hartford (Jul 2025 – Jul 2026), and CVS Health / Aetna before that (2007 – 2025). 20+ years in enterprise software. Open to new roles.
- Connecticut · U.S. citizen
- duong.nick@gmail.com
- linkedin.com/in/newton-duong
- github.com/nduong2000
- duong.casa
Summary
I'm an AI/ML engineer and technical architect with 20+ years in enterprise software, most of it Java and Spring at regulated insurance and healthcare companies. For the last two years I've built production AI agents in Python, along with the systems around the model: the data layer, agent memory and safety controls. I'm a neurodivergent engineer. I think in knowledge graphs, and once a problem has my attention I follow it to the root cause. I'm looking for Staff or Principal roles in AI engineering and AI architecture.
Experience
The Hartford Senior Staff Software Engineer
Jul 2025 – Jul 2026 Hartford, CT
On the Reliability team I built two internal AI platforms on Google ADK and Gemini, from architecture through deployment, and mentored engineers on agent design, testing and AI-assisted development.
- ReliBot. Incident-response assistant with 80+ tools in 16 domains and 12 sub-agents, covering GitHub, Splunk, Dynatrace and security scanning. Engineers reach it from Microsoft Teams, the CLI or a web chat. A model fallback chain with a circuit breaker keeps it answering when one model fails.
- Agent memory. 3-layer auto-recall: L1 core memory, L2 recent conversation, L3 entity-triggered retrieval. Hybrid search runs on SQLite FTS5 and sqlite-vec, weighted 70% vector and 30% BM25. Every tool call goes through a fail-closed allowlist.
- GitHubScanner. AI security reviewer with 6 specialist agents. It covers OWASP/CWE issues with Java and Spring patterns, and enriches dependency CVEs from OSV, NVD and EPSS. Fixes are exported as patches for a developer to review. Nothing is pushed.
- Trust pipeline. LLMs invent findings, so every finding passes deterministic validators before a person sees it. A scan is flagged when more than 5% of its findings fail grounding checks. Verification mode uses 3-level fingerprints, so a finding closes only when the fix is proven.
- Delivery and testing. Keyless deploys from AWS CodePipeline through Workload Identity Federation to Cloud Run, with Terraform and Secret Manager. 3,400+ tests at 86% coverage for GitHubScanner, 2,200+ for ReliBot.
Architecture notes at duong.casa/projects/docs: ReliBot · GitHubScanner
CVS Health / Aetna Staff Software Development Engineer
Jun 2007 – Jul 2025 Hartford, CT
- RAG and GraphRAG for Data Science. Built a GraphRAG vs vector RAG pilot on 43 source documents. Gemini did schema-constrained extraction into Neo4j (1,943 nodes, 4,081 relationships), and a LangGraph + ChromaDB + Vertex AI embeddings twin ran beside it for comparison. Retrieval uses parameterized Cypher templates chosen by intent, so the model never writes free-form queries.
- Evaluation and deployment. Built an A/B dashboard that compared Gemini models on success rate and latency. 2.0 Flash (100% success, 12.2 s median) beat 2.5 Pro preview (32%, 47.2 s) for interactive use and stayed in production. Moving from an always-on VM to scale-to-zero Cloud Run, deployed through GitHub Actions, cut hosting cost about 98%.
- Spring Boot modernization. Led the migration of legacy Spring applications to Spring Boot with OpenRewrite and Claude Code. I wrote reusable refactoring recipes that other departments adopted, and ran the architecture reviews.
- Enterprise platform engineering. Owned the Aetna Enterprise Framework, the company-wide Spring framework other teams built on, plus KickStart reference apps, REST/SOAP APIs and the Docker/Kubernetes/OpenShift platform. Integrated OAuth2, Siteminder SSO and Redis. Mentored project teams and handled production troubleshooting.
- API management and CI/CD. Set API management patterns on CA Layer7 and IBM API Connect, and CI/CD standards on Jenkins, Nexus and UDeploy. Built the Splunk monitoring dashboards.
Architecture notes at duong.casa/projects/docs: GraphRAG vs Vector RAG
Earlier experience
- Pratt & Whitney (UTC), Staff Engineer / Technical Team Lead. Fleet engine diagnostics and forecasting apps on a Spring, Tapestry and Hibernate SOA with Oracle, and flight data transmission with XML digital signatures and encryption.
- Aetna, Senior Java Developer. Enterprise integration middleware on MQSeries and IBM Portal applications. Supervised offshore and junior developers.
Selected personal projects
ClawBot, StockClaw, DVPhoto and the energy platform run on my own homelab. Aquinas is a pre-pilot prototype. Private repos; walkthrough on request. Details: personal projects.Details at duong.casa/projects/personal.
- ClawBot. My agent platform on the Claude Agent SDK: 12 bot instances on Telegram, 220+ in-process MCP tools across 26 domains, a 4-tier Haiku/Sonnet/Opus router, hybrid vector + BM25 memory and 3,500+ tests. It runs my smart home and messaging every day.
- StockClaw. An overnight loop where Claude rewrites a stock-ranking pipeline and a hidden walk-forward backtester decides what stays. About 2,500 candidate rewrites since May 2026, roughly one in ten kept. Research only; it places no orders.
- DVPhoto and DVPhoto for iPhone. Self-hosted family photo platform: a Swift/Vapor + Postgres server with 338 XCTest tests, a React/TypeScript web app and a native Swift 6 backup app in family use through TestFlight. Local face clustering on the Mac mini grouped 24,733 faces from 26,271 photos.
- Energy-market research platform. DuckDB medallion warehouse with 90M+ hourly MISO nodal prices, ported to SPP, PJM and ERCOT, plus PyTorch forecasters. My audit harness caught a look-ahead leak in my own research and flipped 5 of 8 headline backtests negative, so I retired my best-looking number.
- Aquinas. Prototype self-hosted tutor for K–12 schools that asks questions and lets students work their way to the answer. Microsoft 365 SSO, a privacy-first design aimed at COPPA and FERPA, 4 LLM provider adapters and 219 tests. Pre-pilot.
Skills
- Agents & LLMs
- Claude Agent SDK, Google ADK (incl. A2A), MCP, LangGraph, LangChain, Claude (Opus, Sonnet, Haiku), Gemini on Vertex AI, AWS Bedrock, prompt engineering, model A/B evaluation
- Memory & retrieval
- RAG and GraphRAG, Neo4j, ChromaDB, SQLite FTS5 + sqlite-vec, hybrid BM25 + vector search, MMR re-ranking, embeddings
- ML
- PyTorch (LSTM, TCN, PatchTST, GNN), MLflow, time-series forecasting
- Languages
- Python, Java, TypeScript, JavaScript, SQL, Swift
- Frameworks & APIs
- Spring Boot, Spring Framework, FastAPI, React, Node.js, Vapor, SwiftUI, REST/SOAP, microservices
- Cloud
- GCP (Cloud Run, Vertex AI, Secret Manager, GCS, Workload Identity Federation), AWS (Lambda, Step Functions, API Gateway, S3, Bedrock, CodePipeline), Azure
- DevOps
- Docker, Podman, Kubernetes/OpenShift, Terraform, GitHub Actions, Jenkins, Nexus, UDeploy, OpenResty, systemd
- Data
- PostgreSQL, Oracle, DuckDB, SQLite, Redis, Neo4j
- Security & operations
- OSV/NVD/EPSS CVE enrichment, OWASP Top 10 for LLM apps, fail-closed tool policy, prompt-injection fencing, OAuth2/SSO, Splunk, Dynatrace
- Testing & tooling
- pytest, JUnit, XCTest, Playwright, mocked-LLM boundary tests, Claude Code, Codex, GitHub Copilot, OpenRewrite
Education
- M.S., Computer Information Systems, Boston University
- B.S., Information Systems, University of Connecticut