Index

Lessons from shipping a client-ready React 19/TypeScript front end for an AI analyst: SSE streaming with click-to-source citations, constrained tool-calling with stable chart ids, and human accept/reject gates in front of every autofill and export.

Fine-tuning a hierarchy of DeBERTa/RoBERTa-family transformers to route records through a 4-level, 137-path taxonomy — and how class-weighted loss, focal loss, and branch oversampling took the deepest level from roughly 4% to 77%+ accuracy.

Building an offline eval harness that compares frontier and baseline models on latency, token cost, and output quality — with automated hallucination scoring — before any of them touch production.

Provider-native citation grounding as an explainable-AI trust layer: page-anchored source quotes rendered as click-to-source footnotes, claims validated with JSON-LD → RDF and W3C SHACL, and an honest no-evidence state instead of fabricated footnotes.

A tool-RAG discovery surface — five verbs instead of seventy-five schemas — that lets an agent find, understand, and safely compose a large tool catalog without collapsing its ability to pick the right one.

Complete series index: engineering LLM platforms that survive production — agent tool discovery at scale, grounded citations, eval-driven model selection, fine-tuned classifiers under extreme imbalance, and trustworthy AI front ends

The definitive writeup of the alanmargolies88 memorial AI: mining 7,554 recorded matches, reverse-engineering CPS-2 RAM until the bot could truly see, hitting the hard ceiling of imitation, and rebuilding the objective so 'plays like him' lives inside the reward — with every diagram, equation, and honest scoreboard.

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.

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

Reverse engineering MvC1's hidden character box system - discovering three table formats, cursor navigation mechanics, and the path to making Armored Spider-Man selectable

Systematic binary analysis discovers 85 code references, 4 disabled load points, and 3 distinct table formats for character ID 0x2E -- evidence of an unreleased Armored Spider-Man in MvC1's CPS-2 ROM

Pushing every ROM type to the theoretical maximum of Capcom's CPS-2 hardware: 128MB graphics, 16MB audio, 4MB program - documented with in-game verification

Complete series index: reverse engineering Capcom's CPS-2 arcade hardware from first principles -- binary forensics, ROM expansion, audio injection, and character restoration

Engineering a new menu item and 24-line verification screen into CPS-2 arcade firmware using 68000 assembly, jump table extension, and sprite system mastery

Porting MvC2's jazz-fusion soundtrack to MvC1's CPS-2 hardware - from 8-bit PCM injection to multi-bank streaming playback via Z80 ISR-driven bank switching, including the int16_t bug that tripled playback speed

Curated engineering case studies: production GenAI/LLM platform engineering, a memorial fighting-game AI, and binary-level systems engineering on CPS-2 arcade hardware

Designing an emulator platform with Metal GPU rendering, AI opponents via CoreML/PyTorch, massively parallel training, and full developer tooling for fighting games

Discovering undocumented characters, secret input codes, and unused entity IDs through systematic binary forensics on MvC1's CPS-2 ROM

Expanding every ROM type on CPS-2 hardware to its theoretical maximum - 128MB graphics, 16MB audio, 4MB program - using Python automation and binary analysis

Structuring technical analysis so both humans and AI can execute it - from scattered ROM notes to automated validation pipelines with in-game verification

Reverse engineering MvC1's character select screen - discovering cursor navigation tables, decoding entity ID mappings, and mapping three distinct secret character table formats

Adding a new secret character to a 1998 arcade game through ROM patching, 68000 assembly injection, palette engineering, and input sequence state machines