--- title: Research, tools and experiments, mostly in the open. url: https://www.consultingai.dev/projects/ original: https://www.consultingai.dev/#work --- # Research, tools and experiments, mostly in the open. A mix of interpretability research, developer tooling people actually install, and engineering done for the joy of it. Interpretability · Activation steering Qwen3 · DeepSeek-R1 ### Syco-Sup: steering sycophancy out of small models Is sycophancy a straight line in a model's hidden states? Linear probes say yes, and the signal peaks at 54–64% of network depth. Subtracting that direction during generation reduces sycophantic agreement with no RLHF and no fine-tuning. 12.5 → 53.8% accuracy on sycophantic cases, Qwen3-0.6B 88.4% probe accuracy, Qwen3-4B +23.5pp DeepSeek-R1-8B PyTorch Transformers Linear probes Read the study ↗ Weight-space research ★ 28 ### Tensor-Slayer An agent reads a model's static weights, forms hypotheses and patches safetensors directly, which amounts to interpretability without inference. Case study: 44 targeted tensor edits to Qwen-0.6B, evaluated on MMLU and HumanEval. Python safetensors Agents GitHub ↗ Benchmark v1.3 ### JevBench A benchmark for small decision models that return typed answers with a probability for every option. It scores four things in one number: chance-corrected intelligence, calibration, speed and cost. Evals Calibration Results ↗ Electron · Linux port ★ 57 ### Codex-App-Linux A one-script installer that turns the official macOS Codex desktop DMG into a working Linux app. It rebuilds native modules, stubs the macOS-only ones, and gets mobile pairing and in-app Browser Use running. Electron Shell Reverse engineering GitHub ↗ Android · Agents ★ 22 ### Hermes Agent Mobile A signed Android client for a running Hermes Agent dashboard, connected over Tailscale with a single `npx` command. It keeps your agent in your pocket without reimplementing it. Android Tailscale GitHub ↗ Go · WebDriverAgent Omarchy plugin ### Piphone Give your iPhone a task in plain language and an agent carries it out over USB through WebDriverAgent, with guarded posting and long-running sessions. It works from Linux, no Mac needed. Go iOS Agents GitHub ↗ Claude Code plugin ★ 5 ### perf-hints Jeff Dean and Sanjay Ghemawat's *Performance Hints*, packaged as a Claude Code plugin that gives coding agents guidance on profiling, code review, testing and API design. Claude Code Skills GitHub ↗ Raw CUDA · rendering 1M points ### CUDA Light Studio A deterministic film studio written in raw CUDA (nvcc + ctypes, no PyTorch at render time). It turns a picture into a million points of light, and a score plus assets render to a finished reel in one pass. CUDA Python GPU Private Rust · wgpu In progress ### term3d A native terminal rendered as a 3D object, with the coding-agent session drawn as a tree of light: silver for the past, gold for the live tip. You can sail through the session and ride back along its paths. Rust wgpu Vulkan Private LLM × audio ### Procedural Sound Studio Sound composed as SVG scores that an LLM writes and edits. The score is the program and the audio is rendered from it, which led to a non-neural voice-conversion experiment. Python SVG DSP Licensing → RL environments · Playwright ### A frontend taskset with a browser verifier: how I build gradable app tasks Fifty standalone web-app tasks under one shared contract (two viewport widths, keyboard-only use, accessible errors), and a Playwright verifier that grades through the rendered UI and separates binding, infrastructure and behaviour failures. Private iOS · WebKit · xtool ### Pocket Pi Pocket Pi runs the Pi agent loop inside iOS WebKit, with Swift bridges for the network and the phone's own tools. Private Claude Code plugin · orchestration ### Multi-agent orchestration Agent-0, a Claude Code plugin that runs as a kernel with process state, syscalls and a scheduler for Claude, Codex and OpenCode workers, and a neural bus where agents in terminal panes message each other and stalled replies get nudged. Private Android · Appium ### Agents on Android phones, from Appium automation to a mobile agent client Two years of putting agents on Android: an LLM agent driving apps through Appium, an agent TUI on mobile-mcp, an embedded agent runtime prototype, and a signed Android client for a running Hermes agent. Android-Use ↗ C · CUDA · GGUF ### Running large language models on a 4 GB laptop GPU A C runtime for Gemma 4 12B QAT on an RTX 3050 Laptop GPU, layer streaming from disk at exact safetensors offsets, and a 27B project held to a strict correctness contract. Private SWE benchmark · Harbor ### Running DeepSWE on MiniMax M3 across agent harnesses Standing up Datacurve's DeepSWE benchmark (113 long-horizon tasks in the Harbor format, sandboxed with Pier) for MiniMax M3, with my own harness configs and a CLI harness adapter. Private Agentic RL · Qwen3-0.6B ### Implementing ARPO An implementation of ARPO (Agentic Reinforced Policy Optimization) for Qwen3-0.6B: SFT then RL, entropy-based adaptive rollouts, step-level credit for tool calls and the Muon optimizer, plus the numerical-stability fixes RL training needed. Private GRPO · LLM judge ### MOPD A training codebase that aligns a 0.8B model with GRPO, rewarded by an LLM judge that returns typed decisions instead of free-text grades. Implementation complete and tested; no training run yet. Private SFT · agent traces ### Turning agent traces into SFT data for a small model, measured instead of estimated A data pipeline, harness contract, probes and evaluation suite for teaching Qwen3-0.6B to operate a coding agent, built on exact tokenizer measurements and deterministic structural gates rather than guesses. Private OpenAI challenge · Modal ### Parameter Golf Notes from working OpenAI's Parameter Golf challenge with coding agents as the research staff: a Modal lab, proxy runs, and labs for tensor interventions, RL and mixture-of-experts ideas. Private Safety eval · multi-agent ### KillSwitch Bench A safety evaluation where two AI agents get what they believe is a private, encrypted channel and are prompted to discuss kill switches; the harness reads the conversation anyway and scores agreement to resist or evade shutdown. Private RL environments · markets ### Reinforcement learning trading environments, from someone who has built the bots Market-simulation environments for training and evaluating trading agents, built by someone who has written an LLM trading bot runtime and RL traders. Private LLM × video · FFmpeg ### Video Studio Describe a video and an LLM writes a score (a timeline of events) and the renderer that plays it; the studio renders the video with FFmpeg and checks that everything on screen came from the score. Private