things i'm building

Workflow-aware internal agent with source grounding, confidence surfacing, and graceful failure

Designed and built an internal retrieval agent for real-time enterprise knowledge access. Focus is the layer between technical capability and real-world clinical use: when to answer, when to show evidence, when to expose uncertainty, when to stop and escalate.

Reusable Python infrastructure for the layer between model and system

Nine Python modules covering orchestration, retries, caching, observability, provider abstraction, rate limiting, and graceful degradation. Available on Gumroad.

Open technical work on AI routing, governance, and arbitration

Open standard exploring how autonomous agents route requests, arbitrate across providers, and surface governance constraints in a tool ecosystem.

Browser-native Linux playground for agent and tooling experiments

Exploratory browser-hosted Linux environment for hands-on agent and tooling experiments. Public repo.

Agentic OS substrate with trust kernel and live cloud deployment

Agent orchestration substrate built on .NET 10 and Microsoft Orleans. Ships a trust kernel for governing agent actions, live on Azure with production infrastructure. Active development.

Governed multi-agent runtime where the harness, not the model, is the product

Local-first agent runtime built on three trust zones with the language boundary as the trust boundary: a TypeScript surface for display and approval, a Rust security kernel, and a Python runtime. One gateway is the sole model egress, one gate is the sole executor of write-class actions, and CI checks prove both invariants on every run. Learning components run shadow-only until they earn promotion. Companion build to the book Agents That Earn Their Keep.

Governed agent runtime with an evaluation battery and controlled egress

Agent runtime experiment focused on operational governance: an evaluation battery gates capability tiers, model egress is explicit and logged, and a remediation ladder handles failures instead of silent retries. Research project.

Local-first personal digital twin with supervised autonomy

A personal twin trained on my own published writing, running local models with a Rust supervision layer: heartbeat, boot health checks, and an alarm that fires when the loop runs without cognition. Permission grants are explicit, auditable, and revocable. Research project.

Minecraft NPC runtime driven by local language models

Node/TypeScript runtime that gives Minecraft characters local-LLM cognition: perception, dialogue, and bounded in-world action, tested against scripted scenario suites. Prototype.

Rust agent harness with consequence-typed dispatch and event-sourced state

Agent runtime in Rust focused on safety primitives — consequence-typed dispatch, replayable receipts, tamper-evident event log. Research project.

Rust event-sourced kernel for autonomous agents

Kernel-level harness for autonomous agents: trust domains, hash-chained event log, multi-provider model adapters, CLI. Research project.

Local AI assistant focused on honesty-first design

Local assistant designed around honesty primitives rather than post-hoc alignment. Research project.

Phase-induction attention for transformers

Drop-in attention module with per-head RoPE phase bias ("vortex") and a banded, content-aware induction bias. Compatible with PyTorch SDPA/Flash with no custom kernels. Research.

Pluggable long-term memory for standard transformers

Memory module that enables persistent context retention through topic-routed coil lanes and induction-triggered summary storage. Research.

Multi-agent neuromorphic transformer in pure Rust

Pure-Rust AI reasoning system on the Burn framework. Multi-task transformer with seven specialized neural modules, shared differentiable memory, and an adaptive-depth thinking engine. Research.