Home
Personal Background:
I architect, build, and ship production AI systems end to end — scoping the problem with stakeholders, designing the architecture, building and validating it, and owning it in production, often at 10M+-user scale across highly regulated industries. I have led this work since 2015 (data science since 2014), spanning Computer Vision, Natural Language Processing (NLP), and, most recently, Large Language Models (LLMs), multi-agent systems, and Generative AI. I am currently a Principal AI Engineer at FICO, delivering credit-scoring and analytics solutions for financial institutions, consumers, and businesses worldwide. Previously I was a Tech Lead at Vertex Inc, a global leader in tax technology, and a Senior ML Engineer at S&P 500 company LabCorp, building AI across drug diagnostics, drug development, and operations for leaders in the life sciences (see Labcorp SEC filings here). Earlier roles include enterprise Data Scientist at Bayer (a EURO STOXX 50 company), Quantitative Researcher (apprenticeship) at AQR (alternative quantitative strategies and factor-based trading), and Equity Trader at T3 Trading on Wall Street (licensed Series 56 by FINRA). Across these I have partnered directly with customers and executives — leading teams and delivering solo — turning ambiguous business problems into deployed platforms, mentoring the engineers who operate them, and owning the strategy, ROI, and P&L end to end. I think in reusable platforms rather than one-off projects, and I am equally at home in the boardroom and in the codebase. I also run a small algorithmic-trading fund (since 2011) and my own monetized YouTube Channel — feel free to connect on LinkedIn.
Though I began in finance, my AI career grew out of academia. I was a PhD student in Statistics at Columbia University from September 2020 to December 2021, hold a B.A. in Math and an M.S. in Finance from the University of Rochester, and an MBA from the University of Chicago Booth School of Business. My research centers on representation learning — Feature Learning, Deep Learning, Computer Vision (CV), and Natural Language Processing (NLP) — with earlier work in Financial Economics and Asset Pricing. I also teach graduate-level AI/ML and mentor students and early-career engineers, translating research into systems teams can own and run.
Architecture & Infrastructure Patterns:
Across a decade of building and shipping full-stack production systems — from whiteboard to POC to MVP to production, and on to scale at 10M+ users — I operate as a platform engineer at the director / senior-director level: architecting, building, and running cloud-native, event-driven AI infrastructure end to end, solo or as lead. The matrix below maps the stack I apply across regulated industries — frontend through cloud-agnostic deployment — with security and scalability designed in from day one.
| Stack layer | Logistics | Healthcare — Clinical Docs | Banking — Customer Support | Education — AI Tutor | FinTech — Trading Support |
|---|---|---|---|---|---|
| Frontend | SvelteKit / React ops console Live shipment status over WebSockets | React clinician UI Ambient-scribe note review, risk-highlighted | React chat UI Streamed agent replies over WebSockets | SvelteKit tutor UI Streamed feedback + graded results | SvelteKit trading dashboard Real-time quotes & charts streamed to browser |
| Backend | FastAPI · API → queue → worker → Redis Agentic reroute jobs, retries + buffering | FastAPI · audio → transcript → drafted note Async workers + verification pass | FastAPI · Planning → Action → Reviewer loop Dependent tool calls (verify ID first) | FastAPI · deterministic grader + LLM-feedback service Async job queue | FastAPI · low-latency market-data ingest Redis pub/sub → WebSocket fan-out |
| Database | DynamoDB / Firestore / Cosmos DB Vector store over customs rules + SOPs (citable) | Durable record store Vector index over patient record + guidelines | Account & txn store Vector retrieval of fee policy + account terms | Submission + rubric store Course-library vector store | Positions & order store Vector store over filings + research |
| Agent SDK | Claude Agent SDK · fast tier: classify/draft, frontier tier: reroute reasoning RAG-grounded, citable | Tiered models — fast: transcript cleanup, frontier: clinical reasoning Behavior-only fine-tune | Planner + reviewer agents · fast: balance/why-this-fee, frontier: multi-step | Parametric Q&A vs RAG + sequential for “my submission” Deterministic grade, LLM explains why | Fast: quote/status, frontier: trade rationale · RAG over filings Scored vs golden set |
| Agent Tooling & Interop | Two-way MCP (server + client) JWT-scoped tools · entitlement-aware retrieval[1] | MCP tools over EHR Care-team-scoped[1], least-privilege | Session-scoped tools — only the logged-in user's accounts[1] OAuth 2.1 / JWT | MCP tools scoped to this student + this submission[1] | FastMCP streamable-HTTP OAuth 2.1 / JWT-scoped market & account tools |
| Sandbox CPU/GPU | Ephemeral sandboxes (E2B / RunPod) AWS P-series · GCP A2/A3 · Azure ND | In-boundary transcription/inference AWS SageMaker · GCP Vertex · Azure ML · A100 / H100 | Isolated tool-exec sandbox — holds no creds[2] AWS P5 · GCP A3 · Azure NDv5 | On-demand GPU for exam-season bursts · RTX → A100 AWS · GCP · Azure spot GPU | Low-latency multi-GPU inference (H100) Lifecycle: boot → hydrate → drive → sync → reap |
| Data & Storage | Object: S3 / GCS / Azure Blob · Queue: SQS / Pub/Sub / Service Bus Redis for real-time coordination | Encrypted audio/notes: S3 / GCS / Blob · Redis cache Raw-audio retention minimized | Static policy docs: S3 / GCS / Blob · cache static context (not generations) Streams: Kinesis / Pub/Sub / Event Hubs | Course content: S3 / GCS / Blob · Redis cache Grading jobs: Pub/Sub / SQS / Service Bus | Market streams: Kinesis / Pub/Sub / Event Hubs · Redis pub/sub Filings in S3 / GCS / Blob |
| Cloud-Agnostic Deployment | EKS / GKE / AKS · Helm · HPA + Karpenter, spot nodes Istio mTLS · Bedrock / Vertex / Azure OpenAI · IRSA / Workload Identity / Managed Identity[2] | EKS / GKE / AKS · in-tenant BAA model, no-train Per-clinic isolation · regional deploy | EKS / GKE / AKS · ELB / Cloud LB / App Gateway HPA · SOC 2 controls | EKS / GKE / AKS · HPA + cluster autoscaler Multi-region, multi-timezone | EKS / GKE / AKS · low-latency regional · HPA Managed-LLM gateway: Bedrock / Vertex / Azure OpenAI |
| Security | Istio mTLS + JWT / Okta · entitlement-aware retrieval · per-tenant namespaces[1] Encrypt in transit + at rest · secrets isolation[2] | HIPAA / BAA · PHI encrypted, consent before mic · in-tenant no-train model Care-team-scoped, enforced at query time · per-clinic isolation[1] · audit log | GLBA / SOC 2 · isolation enforced at query time by authenticated session identity — never a prompt or eval[1] Expose only the logged-in user's own tools[1] · prompt-injection closed | FERPA + minors · authenticated session Only this student / submission · per-school isolation[1] | SEC / FINRA · MNPI barriers · PII least-privilege Credential brokering — execution envs never hold keys[2] · full audit trail |
| Reliability & Scalability | ~50k shipments/day · limits: orchestration concurrency, vector throughput, API quota Fixes: cache, model routing, async | 10k → millions of visits · limits: model QPS, transcription throughput Fixes: caching, fast-tier routing, async, regional deploy | 10k pilot → 40M customers, 3× launch spike · limits: model + core-banking API quota, retrieval layer Fixes: model tiering, cache static context, async | Exam-season 20× spike, multi-timezone · limits: API quota, model QPS, vector store Fixes: fast-tier routing, async, HPA + load balancer | Market-hours bursts, 10M+ users, low-latency Fixes: model tiering, Redis fan-out, HPA, cache static context |
[1] Multi-tenant data isolation. Systems architected so one customer or user can
never access another's data — per-tenant namespaces, entitlement-aware retrieval, and isolation
enforced at query time by the authenticated session identity (never a prompt value or an eval), so
cross-tenant access is impossible by construction rather than by policy.
[2] Credential isolation in agent tool-calling. User and service credentials never
enter the model context or leak while agents call tools — short-lived IRSA / Workload-Identity-scoped
credentials, secrets isolation, and credential brokering so execution sandboxes hold no keys.
Passion Project:
At leisure, I run W.Y.N. Associates, LLC, a registered legal entity in the state of New York, to pilot and drive for-profit personal passion projects.