Systems thatLearn.Scale.Ship.
- $FF4052 NAME
- Daniel Plas Rivera
- $FF40B0 CALLSIGN
- morphus56k
- $FF4008 UPTIME
- 15+ yrs shipping
- ROLE
- Lead Software Architect · McKinsey
- PREV
- Capital One · PwC
- DOMAIN
- GenAI · Data Eng · Cloud · Systems RE
- 8×LLM
- parallel eval platform
- 99.6%
- infra cost cut
- 64MB
- CPS-2 cap-break
- 18
- technical writeups
/ Capabilities — 15 years across the stack
- 8×LLM
- GENAI PLATFORMS
- 8 models in parallel with automated hallucination detection · sub-100ms pgvector over 100K+ vectors · 25+ McKinsey teams
- 99.6%
- DATA ENGINEERING
- 16.5M daily API calls → sub-million on Kinesis · 99.6% cost cut · Glue / Spark / Snowflake
- IaC
- CLOUD & PLATFORM
- Terraform · KMS · compliance audit trails · Capital One and PwC before McKinsey
- $2E
- SYSTEMS & RE LAB
- 68000 assembly · custom emulator cores · 64MB GFX cap-break · memorial fighting-game AI (VQ-BeT + PPO on MLX)
/ Index — latest entries
The Memorial Bot Learns to Fight Back: From Behavioral Cloning to Style-Constrained Reinforcement Learning
Six weeks of the alanmargolies88 project: why pure imitation plateaued at a passive average, what ~34 verified arXiv papers changed about the design, how a mislabeled RAM byte silently shaped two training runs, and the first RL run where 'plays like Alan' is inside the objective itself.
Cabinet-Faithful Spectating: Why You Can't Just memcpy a Game State Across Architectures
The sequel to the Arcade1Up protocol RE: disassembling libarcade.so's spectate consumer, discovering the cabinet's blind-apply-and-run model, proving verbatim state resume is impossible cross-architecture, and a debugging saga where the biggest bug was my own screenshot tool swapping red and blue.
The Missing State Schema for the Memorial Bot Was a Paper, Not a Trace Session
Closing two long-open items on the alanmargolies88 behavioral clone — state perception and the four player-struct offsets we'd never wired — using a community reverse-engineering reference for Marvel vs. Capcom's EU 0.2.5 ROM.
Can You Bring a Fightcade Player Back? Building a Behavioral Clone of alanmargolies88
An AI-for-good project: training a behavioral clone of a beloved Marvel vs. Capcom 1 Fightcade player from 7,554 of his recorded matches, so the community can spar with his playstyle again.
Reverse Engineering Arcade1Up's Online Fighting Protocol
From TCP packet captures to P2P input exchange at 60fps — reverse engineering Arcade1Up's proprietary online multiplayer system for Marvel vs Capcom, building a desktop spectator client and fighting bridge with RetroArch + FBA Alpha
How Do You Add a New Secret Character to a 1998 Fighting Game?
Reverse engineering MvC1's hidden character box system - discovering three table formats, cursor navigation mechanics, and the path to making Armored Spider-Man selectable
/ Research series
/ Currently on the bench
- 25+
- GENAI AT SCALE
- multi-LLM document-processing platform serving 25+ McKinsey teams in production
- $FF4000
- RL RUN V2
- style-constrained PPO fine-tune of the memorial bot — full matches, behavior-vector shaping live in the loss
- $2A
- DARK SAKURA PORT
- MSHvSF character table ported into MvC1 — framedata byte-identical, 9,439 tiles into the 64MB upper region
- 64MB
- CUSTOM FBNEO CORE
- doubled the CPS-2 GFX address space; upper-region tiles pixel-proven live