Skip to content

Resume

Daniel Plas Rivera — Lead Software Engineer, AI/ML & LLM Platforms. Lead Software Architect at McKinsey & Company, New York; 15+ years shipping production systems. Looking toward LLM infrastructure, agent tooling, model evaluation, and reinforcement learning roles.

2023 — present · McKinsey & Company · Lead Software Architect

AI/ML & LLM Platform Engineering

Lead architect for GenAI/LLM products across 25+ teams, owning AI systems end to end. Built a Model Context Protocol server exposing 75+ agent tools behind a tool-RAG discovery layer, with sandboxed code execution and confirm-guarded destructive ops. Architected a multi-provider LLM gateway (Anthropic, OpenAI/Azure, Google/Vertex, Bedrock) with SSE token streaming and provider-native citation grounding, plus a vision-LLM document-intelligence pipeline with page-anchored citations. Fine-tuned a hierarchical DeBERTa classifier over a 4-level, 137-path taxonomy — deepest-level accuracy 4% → 77%+ via focal loss and branch oversampling — with CI-triggered SageMaker training and ECS Fargate serving. ~120-endpoint FastAPI gateway; React 19/TypeScript front end with streaming chat and click-to-source citations; offline eval harnesses with automated hallucination scoring.

Model Context Protocol · PyTorch · HuggingFace · DeBERTa fine-tuning · LangGraph · RAG · pgvector · SageMaker · FastAPI · React 19 · SSE streaming · evals
2020 — 2023 · McKinsey & Company · Lead Senior Software Engineer

Data Engineering at Scale

Re-architected a Slack analytics platform from 16.5M polled API calls/day to event streaming on Kinesis: sub-million daily calls, 99.6% cost reduction ($47k/yr). Took 10+ Glue/Spark pipelines from 0% to 80% test coverage; closed 10+ critical security findings (fleet-wide IMDSv2, S3 CMK enforcement).

Spark · AWS Glue · Snowflake · Kinesis · Kafka · Airflow · Python · Pydantic · security hardening
2026 · independent research, single-machine MLX

Applied ML Research — Reinforcement Learning & Imitation

A memorial AI that plays like a specific deceased player: VQ-BeT tokenized policy behavior-cloned from thousands of his recorded matches, fine-tuned with style-constrained PPO so 'plays like him' lives inside the objective. Perception: a 266-dim state vector reverse-engineered from arcade RAM. ~34 arXiv papers folded in, every claim adversarially verified — including two audit campaigns against my own training stack. The 2.2M-parameter policy now runs live in the browser via a hand-written forward pass, verified bit-close against the trained model.

PPO · behavioral cloning · VQ-VAE · imitation learning · MLX · reward shaping · browser inference · evaluation protocols
ongoing · self-directed research program

Systems & Reverse Engineering

Multi-year program on 1998 arcade hardware — no source, no docs, encrypted program code. Custom emulator cores that break the 32MB graphics ceiling to 64MB (pixel-proven); a Dreamcast soundtrack running on the original QSound DSP via hand-assembled Z80; a commercial cabinet's netplay protocol recovered from ARM32 disassembly. Every claim terminates in a reproducible artifact.

68000 assembly · Z80 · C++ · Ghidra · binary analysis · emulation · network protocols · Lua instrumentation
/ Earlier — the full arc
PwC
2019 — 2020
Developer (Contract)PwC Data Landscape — client-data security-risk analyzer; Django/Graphene/GraphQL backend for React SPA
Capital One
2019
Master Software Engineer (Contract)re-architected the Credit Risk Rating Platform (probability-of-default microservices) powering credit-approval decisions
Cox Automotive
2018 — 2019
Sr. Python Full-Stack Engineer (Contract)Finance & Insurance platform: Docker/CI-CD infrastructure, performance work via New Relic + Splunk
PwC
2017 — 2018
Developer (Contract)VPN/security tech into the internal tax-audit platform on Azure; Docker-launched VPN containers
Bank of America
2017
Application Architect V (Contract)target-state cloud architecture + platform roadmap; REST APIs, POCs and feasibility studies
Columbia University
2016 — 2017
Application Systems Developertransitioned the university to a Python stack; REST APIs bridging legacy systems, OAuth/LTI integrations
Thinkful
2015 — 2016
Python & Web-Dev Mentormentored career-changers through project-based curriculum, weekly 1:1 code reviews
Earlier era
2009 — 2015
Full-stack lead & founder rolesBasics Plus (lead dev, POS-connected e-commerce), Work-Bench, Brainy Maps (founder), K.I.S. Foundation, NYC College of Technology lab technician
/ Education & certification
AWS Certified Cloud Practitioner
Amazon Web Services
Emerging Media Technologies
New York City College of Technology
Multimedia Programming & Design
Borough of Manhattan Community College
Emerging Media
Hunter College
Cisco Networking + CompTIA PC Repair
School of Cooperative Technical Education
/ Recognition
AngelHack 2016 Manhattan — Grand Prize (Hackcelerator) + Best Use of AWS
New York Minute: fastest/cheapest NYC routing across Uber + public transit (Elastic Beanstalk, EC2, Route 53, S3, Alexa, ArcGIS)
Hack NYU 2014 — MVP Winner
SudoLove: connecting participants, sponsors and mentors for global hackathons