Ken Faiman

Applied-AI Engineer  /  Character & Simulation TA

Agent systems,
and the harnesses that test them.

I build applied-AI agents and the evaluation harnesses that prove they still work. Before that, 25 years of digital-character systems for film and AAA games, from The Matrix to World of Warcraft.

Greater Los Angeles / Orange County Remote-US Primarily AI/ML engineering; also senior technical art
Ken Faiman

Agents, evaluation & open source

At Meta I built a multi-agent deep research system, took it from prototype to a company-shared tool used by 50+ peers, and rebuilt it across 200+ ranked versions against a 15-axis rubric. I also shipped devserver loops that rewrote their own skills from error logs, and a C++ FBX to OpenUSD library that ran as a CI gate so bad geometry could not reach the engine, across hundreds of thousands of assets.

Mainframe MCP

mainframe-mcp is a knowledge server that runs on my own GPU: Qwen3 embeddings, logit-scored reranking, and memory that notices when a new note contradicts an old one. Rebuilding how the reranker scores took hit@1 from 0.30 to 0.40.

Architecture note

Adventure game control

Natural language and voice control for 1980s graphical and text adventures, running on 21,359 lines of my own C++ interpreters. The agent also plays unattended, finishing a full-length adventure at 159 of 159 points with cheats disabled.

Architecture note

Generative type

A FLUX.2 LoRA that draws a full 94-character alphabet from two reference letters, trained on 925 typefaces and vectorized out to OTF and WOFF2.

Architecture note

Infrastructure for agent fleets

vram-mcp arbitrates one GPU across parallel agent sessions so they stop evicting each other's models. hardline-mcp puts Claude Code, Codex and Hermes on one line, so a second model can check the first. Both public.

Architecture note

Sports forecasting

Forecasting over 7 million rows of event data, with quantile models per target and a MILP solver picking lineups against salary and roster constraints. Backtested across 552 nights of real results.

Architecture note

Claude Oracle

claude-oracle fans one research question across chains of ten Haiku scouts, each with its own Sonnet compressor that sees only its chain. Isolation is enforced in Python rather than asked for in a prompt. MIT.

Architecture note

Hermes Realtime

A realtime voice runtime for Hermes Agent. Speech and long-running agent work hold separate cancellation scopes, so cutting off a sentence cancels the sentence and not the task behind it.

Architecture note

MotionTerminal

One Unreal Engine 5.8 Animation Blueprint Template driving four rigs of wildly different proportions, from a 158 cm character to a 198-bone rig at 210 cm, through the IK Retargeter.

Architecture note

gray.bot

gray.bot screens photography by pixel statistics, drafts posts with an LLM, and puts every one in front of a human before it publishes.

Architecture note

Public repositories are on GitHub, where two smaller projects are also published. Private demonstrations are available on request, and Meta agentic pipelines and physics-ML training data remain NDA-bound.

  • reloadedA Windows Terminal launcher for Claude Code that redeploys a whole working arrangement of sessions, repo, window, tab order and screen position, on demand or at logon.
  • sushi-barA Claude Code plugin marketplace: screenshot capture into a session, a richer status line, and two plain-language output styles.

The testing discipline came from simulation

I was Technical Lead of the Avatar Hair team at Meta, Rigging & Simulation Lead for seven years at Blizzard, and lead of a CharacterFX team of 25+ at DreamWorks. The work was systems rather than shots: one rig or solver that has to hold up across a whole slate of characters.

A cloth sim can read perfectly in one frame and be broken in every other. A result can look right and still be wrong. Generative and agentic systems fail the same way. So I build agent systems and retrieval pipelines together with the evaluation harnesses that prove they still work.


Demo reels

Blizzard CinematicsFull character rigging, including facial and muscle, plus cloth, hair & skin simulation.
Feature-Film VFXDreamWorks features & digital-human facial work on The Matrix sequels.

Craft & production

Real-time avatar systems

Avatars & Horizon Meta Reality Labs · 2022–2026

Built core avatar rigging and deformation systems in a VR/mobile engine: auto-rigging, autofit deformation, skinning, blendshapes, skeleton standardization, node-graph clothing-rig automation, and runtime animation & simulation graphs within performance budgets. Avatar Hair Technical Lead. Delivered 100 character rigs for a physics-based ML rigging model.

Rigging & Simulation Lead

Blizzard cinematics Blizzard Entertainment · 2015–2022

Full-body, facial, muscle, clothing and hair rigging across World of Warcraft, Overwatch, Diablo, Heroes of the Storm and Hearthstone. Architected a simulation branch of the studio’s modular graph-based rig-build system with PyQt tooling. Rigging and shot simulation across 25+ cinematics, as a lead.

CharacterFX Lead

Eight feature animations DreamWorks Animation · 2005–2015

Led a CharacterFX team of 25+ and built hero-character cloth, hair and skin across How to Train Your Dragon, Kung Fu Panda, Puss in Boots, Bee Movie, Over the Hedge, Turbo, Penguins of Madagascar and Home. Developed cloth-based skin solvers and core departmental Python tools.

Autonomous crowd agents

The Chronicles of Narnia Rhythm & Hues Studios · 2004–2005

Motion editing, procedural retargeting for autonomous crowd agents, and motion-capture TD work on The Lion, the Witch and the Wardrobe.

Digital-human facial capture

The Matrix sequels ESC Entertainment · 2002–2004

Facial rigging, sculpting, match-move and texture projection on award-winning Universal Capture for Reloaded and Revolutions. Team member and primary contributor to “Making of The Superpunch”, screened at SIGGRAPH 2004 Electronic Theater.


Contact