things i'm building

use it for

• ·Answering policy and procedure questions at the point of work, without digging through source documents
• ·Cutting time-to-answer for frontline staff in the middle of live workflows
• ·Showing the source passage behind every answer so users can verify instead of trust

worth knowing

• ·Explicit "i don't know" behavior instead of confident guessing
• ·Escalation path for questions the system shouldn't answer
• ·Adopted into daily clinical use

use it for

• ·Wiring an LLM provider into an existing product without rebuilding retry, caching, and rate-limit plumbing
• ·Swapping model providers behind one abstraction instead of rewriting call sites
• ·Adding observability to AI calls you already make

worth knowing

• ·Nine composable modules, adopt one or all
• ·Provider-agnostic by design
• ·Failure-first: graceful degradation over silent errors

use it for

• ·Designing how multi-agent systems route requests across competing model providers
• ·Standardizing the governance constraints an agent must surface before acting
• ·Reference material for arbitration between tool providers in an agent ecosystem

worth knowing

• ·Open specification, public repo, provider-neutral
• ·Governance treated as a protocol concern, not an afterthought

use it for

• ·Running Linux tooling experiments straight from a browser tab, nothing to install
• ·Sandboxed agent and CLI experiments with zero blast radius
• ·Teaching and demo environments that work on any machine

worth knowing

• ·WebAssembly-based, runs entirely client-side
• ·Public repo

use it for

• ·Orchestrating fleets of agents with a trust kernel deciding what each one may do
• ·Cloud-native agent workloads on virtual-actor infrastructure
• ·Testing agent permissioning models against real production infrastructure

worth knowing

• ·.NET 10 + Microsoft Orleans virtual actors
• ·Live on Azure with production infrastructure
• ·Trust kernel gates every agent action

use it for

• ·Answering practical questions from a curated knowledge base with zero connectivity
• ·A household preparedness hub that keeps working when infrastructure doesn't
• ·Running local-model tooling in fully disconnected environments

worth knowing

• ·24 integrated tools, 1,000+ automated tests
• ·Everything runs local, models, knowledge, comms, power management
• ·Built on the open Project N.O.M.A.D. foundation

use it for

• ·Coordinating multiple autonomous agents and humans in shared channels
• ·Bridging an agent fleet into one comms fabric instead of point-to-point integrations
• ·Reaching your agents from a phone, web PWA and native Android clients

worth knowing

• ·Protocol-first: the spec is the product, the server is the reference
• ·Live deployment with real agent members, not a demo
• ·Token-authenticated with an audit trail

use it for

• ·One screen of live situational awareness built entirely from public data feeds
• ·Watching real orbital mechanics, Keplerian + J2 propagation over the CelesTrak catalog
• ·Exploring what open data can show: military aircraft detection from public ICAO ranges

worth knowing

• ·25+ live data services integrated, all free/public APIs
• ·Degrades gracefully, works with zero API keys
• ·Visual modes: FLIR thermal, night vision, CRT

use it for

• ·Letting an agent build a UI live in front of you, with pause points you declared in advance
• ·Guaranteeing agent-rendered UI passes redaction and provenance checks before a human sees it
• ·Regenerating only the affected component sub-tree instead of the whole surface

worth knowing

• ·The control plane is the product, the workbench is its first client
• ·Tamper-evident provenance anchored to a transparency log
• ·Working prototype with a governed sprite-generation engine, 85 tests green

use it for

• ·Pair-building software with an agent inside a real, sandboxed Linux environment
• ·Sharing exact live machine state as a URL instead of a "works on my machine" argument
• ·Publishing frozen, content-addressed project capsules others can verify

worth knowing

• ·Agent steps are snapshotted with a replay timeline
• ·Four proven modules composed: sandbox, share, publish, egress
• ·Hardened proxy wiring for controlled network access

use it for

• ·Running agents where every write-class action passes through one auditable gate
• ·Separating display, security, and runtime into distinct trust zones
• ·Companion reference build for the book Agents That Earn Their Keep

worth knowing

• ·Language boundary as trust boundary: TypeScript surface, Rust kernel, Python runtime
• ·CI proves the invariants on every run, one egress, one executor
• ·Learning components run shadow-only until they earn promotion

use it for

• ·Gating agent capability behind demonstrated evaluation results, not vibes
• ·Auditable, explicit model egress for compliance-sensitive environments
• ·Structured failure remediation instead of silent retries

worth knowing

• ·Evaluation battery decides what tier an agent may operate at
• ·Every model call is explicit and logged

use it for

• ·NPCs with real dialogue and bounded in-world behavior, driven by local models
• ·Testing agent cognition in a game world where mistakes are cheap
• ·Scenario-suite evaluation of agent behavior before it touches anything real

worth knowing

• ·Node/TypeScript runtime, local-LLM cognition
• ·Behavior validated against scripted scenario suites

use it for

• ·Agent runtimes where every action carries a typed consequence before it executes
• ·Replayable audit trails for agent decisions, reconstruct exactly what happened and why
• ·Safety-first dispatch as a design primitive, not a wrapper

worth knowing

• ·Rust throughout
• ·Tamper-evident event log with replayable receipts

use it for

• ·A kernel-style foundation for building your own agent stack on top
• ·Multi-provider model adapters behind one interface
• ·Hash-chained event logs where tampering is detectable by construction

worth knowing

• ·Trust domains isolate what each component may touch
• ·CLI-first workflow

use it for

• ·A local assistant that says "i don't know" by design instead of guessing
• ·Studying honesty as a first-class architectural primitive rather than a fine-tuning patch

worth knowing

• ·Local-first, your data stays on your machine
• ·Honesty primitives built into the design, not bolted on after

use it for

• ·Drop-in attention experiments on existing PyTorch models
• ·Exploring induction-bias improvements without writing custom CUDA kernels

worth knowing

• ·Compatible with PyTorch SDPA and Flash attention
• ·Per-head phase bias, swaps in without architecture surgery

use it for

• ·Adding persistent long-term memory to a standard transformer
• ·Topic-routed retention without reprocessing the full context every time

worth knowing

• ·Pluggable, designed to attach to existing architectures
• ·Summary storage triggered by induction patterns, not manual curation

use it for

• ·Exploring verifier-style scoring for theorem-proving tactics
• ·Testing Energy-Based Transformer claims on a concrete formal-methods task

worth knowing

• ·Builds on the public EBT codebase with full attribution
• ·Lean 4 proof tactics as the evaluation domain

use it for

• ·Exploring modular neural architectures in pure Rust
• ·ML experimentation without a Python dependency chain

worth knowing

• ·Built on the Burn framework, no Python anywhere
• ·Seven specialized neural modules over shared differentiable memory
• ·Adaptive-depth thinking engine