Local AI operating layers
Model routing, profiles, fallbacks, context budgets, streaming normalization, JSON repair, and app-specific AI lanes.
Useful AI needs a system layer
I build the missing system layer around useful AI.
I build practical AI systems that move beyond demos: local and cloud model routing, agent orchestration, long-context memory, evaluation harnesses, workflow automation, and usable product interfaces.
The work sits between frontier model capability and usable software: app-specific AI lanes, memory infrastructure, multimodal interfaces, and evidence loops that keep humans in control.
What I build
I am not building isolated AI demos. I build the layer around models: routing, memory, workflow state, evaluation, and interfaces that make AI useful inside real product constraints.
Domain-specific AI assistants
ChefAI, Lexi, Friday, and other systems that wrap models in product logic, workflow context, and evaluation.
Durable memory and evidence
Muninn-style memory infrastructure with provenance, promotion and decay, retrieval, deletion boundaries, and traceable context.
Human-in-the-loop interfaces
Sketch and canvas tools, operator panels, workflow controls, and UI patterns that let humans steer AI instead of just chatting at it.
Featured systems
The strongest systems lead the page because they show the current thesis in working form: model routing, memory, evaluation, domain assistants, and human-directed interfaces.
Friday AI + Local LLM Gateway
A local-first AI operating layer for routing models, tools, memory, and workflows across multiple applications.
- Universal gateway and model bridge for app-specific aliases such as friday, chef, lexi, friday-coder, and friday-fast.
- Supports routing, streaming normalization, context budgeting, JSON repair, fallback behavior, and app-specific profile control.
- Foundation layer for Friday, ChefAI, Lexi, and future agent systems.
ChefAI
A production-style culinary AI assistant with structured reasoning, fallback handling, UI guidance, and regression-tested food reasoning.
- Intent router: 35 passing tests.
- Guarded orchestrator web integration: 79 passing tests.
- Knowledge graph pressure suite: 15 passing tests.
- Food reasoning sweep: 18/18 pass, 0 partial/fail/error.
- Public demo stabilization completed.
Muninn
A memory prosthetic system for reconstructing, promoting, decaying, retrieving, and visualizing useful long-term context.
- Muninn is not an archive. It is a memory compiler.
- Uses evidence, provenance, promotion and decay, conflict and supersession, and retrieval.
- Designed as shared substrate for Friday, Lexi, ChefAI, and future systems.
SketchMath
A canvas-first sketch-to-CAD system for geometric reasoning, constraints, and mechanical design workflows.
- 2D sketch canvas with constraints, dimensions, selection history, solver status, and closed profile detection.
- CAD-style extrusion and STEP export direction.
- Shows that useful AI interfaces are not always chat boxes.
Lexi
An app-specific AI interface focused on stable identity, clean interaction boundaries, and gateway-backed model behavior.
- Gateway alias integration for app-specific model behavior.
- Stable identity and interaction boundaries.
- Scaffold stripping and clean response boundaries where the runtime supports them.
Local AI Lab / Model Infrastructure
A self-hosted model infrastructure lab for quantization, benchmarking, routing, and long-context inference.
- Local multi-GPU inference infrastructure for open-weight model testing and routing.
- GGUF and llama.cpp, vLLM, quantized deployments, long-context probes, and JSON/coding/streaming evals where supported.
- Purpose: reduce hosted token spend, increase privacy and control, and make app-specific model routing practical.
Built from the operator side.
My background is healthcare, operations, and hands-on system building, not a traditional software ladder. That changes how I build.
I care less about flashy demos and more about whether an AI system survives contact with real users, messy workflows, changing context, constrained infrastructure, and safety boundaries.
How I think about AI systemsWhere I fit
I fit best where the hard part is not calling a model. The hard part is deciding what context it should see, what tools it can use, how it should fail, how humans stay in control, and how the system proves it is getting better.
- AI systems engineer
- AI solutions architect
- AI enablement lead
- Forward-deployed AI builder
- Agent workflow / model routing / memory systems builder
- Human-in-the-loop AI product builder
Full project index
The index stays here as supporting evidence. The current systems are first; older or smaller projects remain available for range and history.
Friday AI + Local LLM Gateway
ActiveInfrastructureLocal AI
Anchor operating layer for model routing, tool lanes, memory, fallback behavior, and app-specific profiles.
ChefAI
DemoEvaluationDomain assistant
Culinary assistant with structured reasoning, fallback handling, UI guidance, and regression-tested food reasoning.
Muninn
MemoryInfrastructure
Memory compiler for evidence-backed context, provenance, promotion/decay, conflict handling, and retrieval.
SketchMath
CanvasCADHuman-in-the-loop
Canvas-first spatial and CAD reasoning surface for constraints, dimensions, solver status, and mechanical workflows.
Lexi
ActiveGatewayInterface
App-specific AI interface with stable identity, gateway-backed model behavior, and clean response boundaries.
Local AI Lab / Model Infrastructure
Local AIvLLMBenchmarks
Self-hosted inference lab for open-weight model testing, quantization, routing, long-context probes, and evals.
Core
Bifrost
Access on request
Cross-project atlas that defines ownership boundaries, capability contracts, terminology, and overlap-risk closure across the ecosystem.
Assistant orchestration proving ground where routed lanes, memory providers, and failure handling are exercised under runtime constraints.
LAILA
Access on request
Next-generation assistant runtime unifying agency, Muninn memory, and identity continuity.
lex
Access on request
Client-facing local-first companion runtime centered on identity continuity, persona control, and avatar-aware interaction.
Mimir
Access on request
Repository self-model and world-state orientation system that brokers bounded rehydration for coding agents across evolving codebases.
Model-agnostic memory substrate with card, evidence, and policy-state primitives plus deterministic rehydration for agent runtimes.
OnHand
Access on request
Inventory and intent platform spanning Android client workflows, enterprise APIs, and optional vision evaluation services.
Commissioned applied competency project for modular trading research and policy-driven execution planning.
Experimental
Althing
Access on request
Profile-aware specialist-routing harness for multi-lane inference experiments with explicit degraded modes and measurable control profiles.
Coding arena platform for challenge publishing, controlled execution, and repeatable submission scoring.
Commissioned applied competency project for deterministic strategy evaluation and replayable market decisions.
Local-first autonomous media pipeline that converts news signals into scripted, rendered, and packaged short-form broadcasts.
Automated playtesting system built specifically to exercise SubSim through deterministic simulation runs.
Sindri
Access on request
Bounded capability-forging service focused on trust-calibrated wrapper lanes, verifier-backed review bundles, and promotion gating.
Audio-first deterministic submarine simulation designed for fast iteration, replay traces, and headless testing.
Commission
Accountant
Access on request
Commissioned applied competency project for secure accounting ingestion, OCR review, and export workflows.
Show older work / project history map
A quieter topology view for readers who want to see how the archive and adjacent systems relate beneath the current story.
-
Assistant orchestration proving ground where routed lanes, memory providers, and failure handling are exercised under runtime constraints.
Client and identity surface
-
Client-facing local-first companion runtime centered on identity continuity, persona control, and avatar-aware interaction.
Routing and lane instrumentation
-
Profile-aware specialist routing with constrained research lanes and non-authoritative shadow characterization.
Profile-aware specialist-routing harness for multi-lane inference experiments with explicit degraded modes and measurable control profiles.
Experimentation waves
-
Audio-first deterministic submarine simulation designed for fast iteration, replay traces, and headless testing.
Automated playtesting
-
Automated playtesting system built specifically to exercise SubSim through deterministic simulation runs.
-
-
Local-first autonomous media pipeline that converts news signals into scripted, rendered, and packaged short-form broadcasts.
-
Coding arena platform for challenge publishing, controlled execution, and repeatable submission scoring.
Commission work (applied competencies)
-
Commissioned applied competency project for secure accounting ingestion, OCR review, and export workflows.
Inventory/vision exploration
-
Inventory and intent platform spanning Android client workflows, enterprise APIs, and optional vision evaluation services.
-
-
Commissioned applied competency project for modular trading research and policy-driven execution planning.
-
Commissioned applied competency project for deterministic strategy evaluation and replayable market decisions.
-
-
Started as embedded memory experiments inside FRIDAY/Lex.
Model-agnostic memory substrate with card, evidence, and policy-state primitives plus deterministic rehydration for agent runtimes.
Now a standalone memory substrate for cards, evidence, policy-state, and deterministic rehydration.
Repository orientation branch
-
Extends memory work into repository world-state orientation, brokered context, and bounded cross-repo export.
Repository self-model and world-state orientation layer with deterministic brokered rehydration and cross-repo validation across real engineering repos.
Ecosystem atlas and boundary governance
-
Cross-project atlas that defines ownership boundaries, capability contracts, terminology, and overlap-risk closure across the ecosystem.
Bounded capability proving lanes
-
Wrapper lane graduated; adjacent-lane transfer proving remains intentionally narrow and safety-gated.
Bounded capability-forging service focused on trust-calibrated wrapper lanes, verifier-backed review bundles, and promotion gating.
-
-
-
Next-generation assistant runtime unifying agency, Muninn memory, and identity continuity.
Interested in working together?
Whether you want help designing a trustworthy AI system layer, pressure-testing an architecture direction, or exploring Mímir as a foundational component, I am open to serious conversations.