An agentic framework
Every project in this portfolio involves AI. This one is where the AI engineering itself is the work. Closed source, research track. The point is a worker substrate an LLM cannot silently break, not a product to ship.
Heavily inspired by what is moving in the industry. I pull references daily, follow new models, frameworks, and architectures as they land, and adapt the pieces that hold up into this framework.
2026 · private · Python · TypeScript · SQLite · Gemini · OpenAI
Numbers from the live runtime, not from the README.
130k
LOC Python, core runtime
12k
LOC TypeScript, tooling and UI
10
distinct worker types: planner, exec, dev, flash, memory, monitor, review, rollback, ui, web
8
dynamic registries, workers, tools, models, providers, executors, verifiers, services, input sources
1,773
production runs already through the supervisor (live in state.sqlite3)
13,960
append-only events in the JSONL log; replayable from any checkpoint
Architecture, what runs vs what is deferred
The system is designed in three layers. Two of them are intentionally empty.
Layer 3, environment. Workers, registries, event log, evidence, supervision. Fully implemented and running today. This is the layer that has to be solid before anything above it makes sense.
Layer 2, cognitive. Memory, goals, self-edit, reflective mechanisms. Partial: full-text memory chunks and embedding cache exist; goals and self-edit are scaffolded but not yet wired into the planner.
Layer 1, sensory. Watchers, ambient context, autonomous ingress. Architecture drafted, no implementation yet. It is supposed to be empty until layer 3 holds.
Event-sourced execution
Append-only JSONL log paired with a 17-table SQLite store. Any run can be resumed from a checkpoint. Tier-1 events (tool.observed, marks, run.*) never drop, even under backpressure.
Policy-gated tool calling
Tools live behind a policy engine: blocklist, allowlist, per-category approval requirements, schema validation on args. UI mutations need explicit approval. Denials propagate; nothing side-effects silently.
Dynamic registries everywhere
Workers, tools, models, LLM providers, executors, verifiers, services, input sources, eight registries, all hot-reloadable from config. Worker types ship with descriptors that declare allowed tools, checkpoints, and binding families.
Backpressure and fail-closed semantics
NeuroBus pub/sub with bounded queues; oxygen system reduces background-task budget when vitals indicate saturation. If an LLM endpoint goes down, the event log gets an llm_unavailable entry and the planner can react. No silent fallback.
From the workspace
The system spends as much time on whiteboards as in code. Diagrams next to event log views, worker dashboards next to architecture sketches.
Also in these roles
AI engineer is the primary lens, by a wide margin. The other roles still apply because the runtime around the model is half the work.
System builder
Most of the 130k LOC is not the model interface, it is the substrate. Orchestration of 10 worker types with restarts, FIFO for the UI worker, defensive scaling. 8 dynamic registries (workers, tools, models, providers, executors, verifiers, services, input sources) hot-reload from config. NeuroBus pub/sub with bounded queues and tier-1 vs tier-2 backpressure. Oxygen system reduces background-task budget when vitals saturate.
Fullstack
Python runtime (130k LOC) plus TypeScript tooling and UI (12k LOC). Same invariants hold across the stack: checkpoints persist, observations index by tool_call_id, marks unblock waiting workers, event log replayable.
Founder
Research track but founder-owned. Every architectural call has to survive the next concrete pressure test (real worker runs, real bench workloads), not the next funding round.
Source private. On a serious research conversation I can walk a real run end-to-end, planner output, worker checkpoints, observations, verifier marks, under a discussion-level NDA.
