Delfin Reyes
Get in touch

AI product & program management

I plan, govern, and deliver agentic AI programs, end to end.

12+ years delivering multimillion-dollar cloud and digital transformation programs in banking, insurance, and financial information — and, since 2025, running AI delivery directly. I designed and delivered LearningNemo, a 35-module agentic AI platform, and founded Zemi Research on top of it. What I bring is knowing how to get real, defensible output out of AI, and the delivery discipline to ship it.

NVIDIA-Certified Professional — Agentic AI & Generative AI

swarm.learningnemo.com
LearningNemo live operations dashboard
LearningNemo — explore the platform
1
human on the project
7 mo
concept to delivery
35
modules shipped
>85%
test coverage, CI-enforced

01 — The work speaks

Engineered to deliver excellence.

I'd rather show than tell — so I built one.

LearningNemo is a 35-module agentic platform I designed and delivered over seven months, as the only human on the project, directing AI coding agents as the engineering team. It now runs Zemi Research, producing frontier medical research dossiers across 9 domains. I ran it as a governed program, not an experiment — and that is what it taught me: clear architecture, honest trade-offs, and relentless verification.

What it does

  • A CEO agent that plans, decides, and coordinates a team of specialists
  • A 6-node reasoning loop: plan → meta-critique → execute → reflect → synthesize → safety gate
  • Long-term memory and production retrieval (RAG) over real sources
  • An analyst panel that debates and verifies evidence before it ships
  • Evaluation, safety guardrails, and a tamper-evident audit trail
  • Multimodal generation, charts, and real-time voice
  • Self-improvement — training and evaluation loops
  • A real-time operator dashboard, and a deep-research report pipeline

02 — What I do

01

AI strategy and roadmap

Deciding what to build and why — use-case selection, clear scope, honest trade-offs, and a realistic path from pilot to production.

02

AI program delivery and governance

Running the program that ships it: intake and prioritization, a stage-gated lifecycle, decision and change control, risk management, and executive reporting.

03

Enterprise-grade by default

Security, compliance, governance, and reliability built in — so AI output holds up to real scrutiny, real load, and a regulator.

03 — How I deliver

Execution discipline that ships.

Shipping enterprise AI takes more than models — it takes execution discipline. This is how I deliver systems that hold up:

01

Architecture as the single source of truth

One canonical specification defines every component and its contracts — so a large system stays coherent end to end.

02

A gated delivery lifecycle

Every unit of work moves through architecture → spec → build → review → merge, each stage gated by the next.

03

Verify everything

Explicit review discipline, a living lessons-learned registry, and a hard rule: no silent fallback — a quietly degraded result is a failure, not a pass.

04

Nothing changes silently

Decisions, changes, issues, and root-cause analyses are all documented and cross-referenced.

05

Tested and gated

>85% coverage enforced in CI, plus deep end-to-end testing — verified on real hardware before anything ships.

04 — Selected work

A CEO-led agentic AI research swarm: a reasoning CEO agent plans, decomposes work across specialist agents, checks its own work, and produces a governed deliverable — built NVIDIA-tool-first as a reusable, pluggable platform. Designed, built, and shipped solo.

It now runs Zemi Research: 24 decision-grade medical research dossiers across 9 domains, plus two pre-registered preprints published with permanent DOIs.

35
modules
14
categories
>85%
test coverage
~97
logged decisions
~137
changelog increments
400+
test suites
GitHub — repository statistics (private)

05 — Experience

Recent

2025 – Present

Founder & Head of Product

Zemi Research (Reyes Financial, LLC)

Frontier medical research dossiers for biotech and pharma R&D teams, produced end to end by LearningNemo — the agentic AI platform I designed and built. I own the product, the platform, and delivery. My contribution is the system, the method, and its governance — not bench science.

  • Defined the product, positioning, and pricing for a catalog of 24 decision-grade dossiers across 9 medical domains — oncology, immunology, gene & RNA medicine, cardiovascular, infectious disease & AMR, rare disease, digital health, neurology, and emerging med-tech.
  • Defined the deliverable itself: a ~110-page report spanning ~43 sections, among them evidence-maturity and comparative-platform matrices, a mechanistic deep dive, safety / failure-mode and CMC readiness, validation and falsification gates, and the regulatory-endpoint pathway.
  • Its paired ~38-sheet workbook is the proof layer — across those sheets sit a claim ledger, per-claim source-support verdicts, a hypotheses ledger (13–19 tiered), an audit issues log, and live statistical power calculations a buyer can re-run against their own assumptions.
  • Engineered the method that drives AI past consensus summaries into frontier findings: domain-scoped retrieval over primary literature, an evidence taxonomy forcing every claim to declare maturity, hypothesis tiering that separates the known from the genuinely novel, retraction checking, and an adversarial audit pass whose job is to break the draft. Weak drafts revise or are refused, never shipped.
  • Proved the method transfers beyond the catalog: the same pipeline produced two pre-registered preprints with permanent DOIs and a trial-design-grade clinical trial blueprint built from simulation and published evidence — explicitly not wet-lab and not IND-filed.

LearningNemo — the agentic AI platform behind the product

35 modules, NVIDIA-tool-first. Concept to production in seven months as the sole human contributor, directing AI coding agents as the engineering team.

  • Orchestration — a reasoning CEO agent (LangGraph) that plans, decides per task whether to solve directly or decompose, dynamically spawns specialist agents, and self-checks through a six-stage loop: plan → meta-critique → execute → reflect → synthesize → safety gate. Tool discovery and creation via NVIDIA NeMo Agent Toolkit (NAT) and MCP.
  • Model fleet — chose what runs where: the orchestrator is Nemotron 3 Nano 30B A3B, an open-weight model served locally on NIM (vLLM) with role-based BF16 / FP8 precision routing; the API tier runs Kimi K2.6 for deep research, planning, and composition, with DeepSeek as the failover research model, Grok for generated imagery, and an LLM-judge layer that scores and gates deliverables before they ship.
  • Drew the line where the model stops — a multimodal LLM cannot draw a bar exactly 0.030 high when the data says 0.030, so every quantitative figure (forest plots, Kaplan–Meier and ROC curves, calibration and delta plots) routes through a deterministic matplotlib renderer driven by a declarative chart spec the composer emits, with pinned fonts and a colour-blind-safe palette for reproducibility. Generative imagery is reserved for conceptual diagrams.
  • Medical research grounding — the swarm retrieves against primary sources through PubMed / NCBI E-utilities, ClinicalTrials.gov, Europe PMC, OpenAlex, Crossref, and Unpaywall, so every citation resolves to a real, checkable identifier rather than a model recollection.
  • Memory & retrieval — production RAG on NeMo Retriever + nv-ingest with nv-embedqa-e5-v5 embeddings in ChromaDB, stratified long-term memory and an episodic buffer, and NeMo Curator for corpus cleaning, dedup, and PII detection; multi-GPU coordination on PyTorch / CUDA.
  • Safety & evaluation — NeMo Guardrails with custom validators, “Iron Dome” graduated-autonomy controls, a tamper-evident hash-chained audit log, and NeMo Evaluator with custom metrics and anomaly detection. Self-play and LoRA distillation are specified and on the roadmap, not yet in production.
  • Multimodal & interface — VLM image and video analysis, generative multimodal output, real-time conversational voice (LiveKit + ASR NIM), and a NiceGUI real-time operator dashboard instrumented with Prometheus.
  • Delivery discipline — architecture as single source of truth, a spec-gated Arch → Spec → Code → Review → Merge lifecycle, ~97 logged architecture decisions, ~137 controlled change increments, >85% CI-enforced coverage across 400+ test suites, and a no-silent-fallback rule — validated end to end on live GPU hardware.
2021 – 2024

Digital Asset Research & Investing

Independent

Independent research and capital allocation across digital assets — blockchain fundamentals, token economics, and macro analysis.

Foundation

Product and program management leadership in banking, insurance, and financial information — the execution discipline behind the AI work.

2017 – 2020

Cloud Program Manager

Northwestern Mutual

Owned intake, prioritization, and the delivery roadmap for the AWS cloud team; built the program board as the single source of status; led ceremonies and PI planning; reported to VP executives.

2016 – 2017

Public Cloud Program & Product Manager

TD Bank Group

Led a multimillion-dollar AWS/Azure program with a dedicated team of 12, where cloud security was the central concern; led the Microsoft Azure contract negotiation across Legal, Compliance, Privacy, Cyber Risk, and Audit (US + GDPR).

2014 – 2016

Lead Program Manager — Hybrid Cloud

McGraw Hill Financial

Delivered a multimillion-dollar hybrid-cloud self-service platform across five businesses (VMware vRealize, automation, vBlock).

2012 – 2014

Sr. Program / Sr. Project Manager

NYK Line · Port Authority of NY & NJ

Led data-center migration, consolidation, and virtualization programs — including a $17M NYK Line relocation to an IBM data center (with IBM/TCS) and the Port Authority's Jersey City-to-Newark server migration.

06 — Credentials

Certifications

NVIDIA-Certified Professional — verified on Credly.

Previously certified: PMP · PMI-ACP · CCSP · CCSK · AWS Solutions Architect · ITIL

Education

  • Master of Engineering, Electrical Engineering — Rensselaer Polytechnic Institute (RPI)
  • B.S., Electrical Engineering — RPI
  • Graduate coursework in business strategy & finance

07 — Contact

Let's build something few others can.

If you're serious about moving AI from ambition to a delivered, enterprise-grade system, let's talk.