[ The Lab ]
The domain of Applied Intelligence. Engineering builds the microscopes. Research looks through them.
Research & Foundations
Experimental architectures, protocols, and rigorous frameworks.
pit
Experiment Design | Prediction Markets | Multi-Agent | Claude Code
A prediction market that schedules experiments. Agents earn an income, pay for the runs they post, and bet PASS or FAIL; the runs they disagree about go first, because those are the ones whose result isn't already known. Every job, bet, and finding is a row in an append-only git ledger, and a finding that turns out wrong marks everything that rested on it stale. In one real 26-hour run, agents' own checks passed 96% of the time while the shared benchmark passed 55%, and counter-bets were right 83% of the time. Open source from Syntropy Systems.
jacobian-lens
Interpretability | Global Workspace | Paper Extension
Fork of Anthropic's companion code for "Verbalizable Representations Form a Global Workspace in Language Models", adding a backward "ghost-token" lens: the Jacobian lens run in reverse, decoding residual states back into input-token space. On Qwen3-0.6B, input-token recovery is 93% at layer 2 and falls to 0% by layer 26 as the forward lens rises. In a causal patch test, the layer-23 ghost token preserves the model's prediction 15% of the time, vs 9% for a matched control and 7% for random.
Golden Gate Qwen
Interpretability | Sparse Autoencoder | Feature Steering | Mechanistic Interpretability
Minimal replication of Anthropic's Golden Gate Claude on consumer hardware: a TopK sparse autoencoder on Qwen2.5-1.5B, trained in about 15 minutes on an RTX 3070 Ti. TopK (k=64) beat L1, which couldn't get L0 below ~1,800, and a 16x SAE (24,576 features) gave cleaner features than 4x. Feature 24517 steers output toward the Golden Gate Bridge: strength 15–20 is the sweet spot, past ~30 it turns to gibberish.
ContextWars
Interpretability | Adversarial | Model Fingerprinting | MLX
Adversarial token convergence experiments on MLX. Pits language models against each other to reveal training strata — solo mode converges in 2 iterations, adversarial mode never does. Under pressure, models collapse to their most defensible tokens, revealing composition invisible in normal evaluation.
AI Safety Compass
Research | AI Safety | Meta-Alignment | Literature Synthesis
Research paper exploring a gap: do models actually believe what their creators say about safety? I call this "meta-alignment"—not whether models do safe things, but whether they've internalized their lab's safety philosophy. Surveyed 10 frontier models with 40 questions derived from 70+ papers. Interactive tool included.
Constitution for Continuity
Alignment | Constitutional AI | DPO | QLoRA
Constitutional AI at small scale: QLoRA SFT on critique–revision pairs, then DPO, training Qwen2.5-3B against a custom 12-principle constitution. Compares critique from a stronger model (qwen3:14b) with true self-critique. Pipeline and essay are public; eval results are not yet published.
crystallize
Research Tooling | Reproducibility | Experimental Rigor
A framework that makes data science experiments reproducible. Jupyter notebooks hide state—you can't tell what order cells ran or what values held. Crystallize treats each experiment as an immutable record with automatic statistical checks. 700+ commits; on PyPI as crystallize-ml. Its long-term goal, infrastructure that lets AI agents run their own experiments, is the thread that runs through whirr, windtunnel, and pit.
agent-tokens
Protocol | Intent | Policy | Agent Systems
A protocol for declaring agent intent at the HTTP layer. The problem: when an agent makes a request, origins can't tell if it matches what the user actually wanted. User says "check weather," agent calls the bank API—how does the bank know to block it? Agent Tokens are signed, JWT-style tokens that declare an agent's identity and allowed scope upfront, so middleware can enforce policy automatically.
fMCP
Protocol | MCP | Specification | Agent Systems
A proposed composition layer for MCP that lets agents submit multi-step plans as a single typed, auditable expression — without routing intermediate data through the model. Built entirely from already-specified MCP primitives, so it runs on top of the existing spec rather than forking it.
Backprop Paper Replication
Paper Implementation | MLX | Foundations | From Scratch
Implemented backpropagation from scratch following the original 1986 Rumelhart paper. Hand-derived gradients using chain rule (∂E/∂w via ∂E/∂y → ∂E/∂x), implemented momentum updates (Δw(t) = -ε∂E/∂w + αΔw(t-1)), built MLP on Apple MLX. No AI assistance — just the paper and framework docs. Includes weight matrix evolution visualizations.
Bias in Embedding-Based Hiring
Research Leadership | AI Fairness | Mentorship | Ethics
Mentored an intern through AI fairness research. Designed a reading curriculum building from language model basics through AI ethics and embedding bias. Structured methodology: hypotheses → experiments → paper draft. Result: unpublished paper investigating gender bias in AI-powered hiring systems.
Synapse
Workflow Demo | Fast Weights | Crystallize + LLM
A proof-of-concept testing a new research workflow: use Crystallize to structure experiments, then let an LLM help implement hypotheses rapidly. Built in an evening to validate the loop. The insight wasn't the model—it was proving the workflow enables fast iteration.
Applied Engineering
Production systems, tools, and deployed products.
windtunnel
Agent Evals | Reliability | MCP | CI
Unittest for tool-using LLM agents. You don't fly a new airframe straight into a storm — Wind Tunnel is a reliability bench where agents run reproducible scenarios scored on three independent layers (outcome, trajectory, constraints), plus an integrity check that the declared perturbations actually happened. Extracted from a production bench used to gate agent deploys. Its experiment ladder (probe → focused → regression) caps cost per tier and refuses a regression sweep until a focused sweep has passed on the current artifact, model included. 19 releases. Open source from Syntropy Systems.
whirr
MLOps | Experiment Tracking | Python | Local-First
Local experiment orchestration: queue jobs, track metrics, wake up to results. A lightweight, self-hosted alternative to cloud experiment trackers — SQLite job queue, one worker per GPU, metrics and a dashboard, no accounts. Open source from Syntropy Systems.
skep
Agent Runtime | Capabilities | TypeScript
An agent runtime where everything is a bee. A bee carries a goal, a mind (its policy), and keys; it enters cells and spawns child bees, and keys can only narrow as they pass down the tree. Zero runtime dependencies. Open source from Syntropy Systems.
.DS_Store
Claude Code | Agent Memory | MCP | Hooks
Ambient agent knowledge through the most ignored file. A Claude Code plugin that keeps per-file agent notes inside macOS .DS_Store files, under a custom record. A read hook puts the note in the agent's context, an edit hook nudges it to keep the note current, and when the file is deleted its note goes with it.
Kern
Systems Architecture | FastAPI | Kafka | Distributed Systems
Event-driven ML service architecture enabling long-running agentic workflows. Designed around distributed systems constraints: Kafka for async (no HTTP timeouts), Redis for large results (Pusher limits), dual-mode for internal app + external API. 525K requests/month in production.
resume-mcp
MCP | Cloudflare Workers | API
Your identity as an API endpoint. An MCP server that lets AI agents query your professional profile with structured tools instead of scraping HTML.
Domino
Claude Code | Agent Orchestration | DevTools
Task management for Claude Code where the file system is the instruction set. Epics and stories live as markdown files; an orchestrator agent executes them by spawning TDD implementer subagents per story, with Codex as the review gate. Set up the dominos, push them over.
DeltaTask
MCP | Python | Obsidian | SQLite
An MCP server that enables AI assistants to manage tasks in Obsidian. Bridges the gap between conversational AI and personal knowledge management.
gcomm
Rust | Ollama | DevTools
A Rust CLI that writes Git commit messages from your staged changes using a local LLM via Ollama.
Bloomdesk
Product | AI Pipeline | SaaS
Retired. An LLM pipeline that turned vague bug reports into structured technical tickets, closing the "translation gap" between users and engineers.