Local compute. Deterministic control. Verifiable outcomes.
Scott Hardie · Solutions Architect · AI Systems Builder · Toronto, Canada
Explore the systems · Read the architecture · View the storefront · Connect
LOCAL-FIRST AI · CONTROL PLANES · DETERMINISTIC BACKENDS · FINOPS · VERIFICATION
AI demos are easy. Production systems must survive retries, partial failure, hostile inputs, runaway spend, model drift, and an auditor asking exactly what happened.
Hardonia is a working portfolio of control planes, runtimes, security boundaries, and commercial systems designed around one idea:
Intelligence can be probabilistic. Infrastructure cannot.
Best-fit work: production AI and platform architecture, technical due diligence, and modernization of systems where security, reliability, or financial correctness matters. I take on selective independent engagements alongside my full-time role.
| Observe | Control | Prove |
|---|---|---|
| Capture agent, model, tool, cost, and transaction events. | Route workloads, enforce policy, isolate tenants, and recover safely. | Replay decisions, verify state, reconcile money, and export evidence. |
Selected systems Platform Evidence All projects Products Contact
The system map above is the visual overview. This table defines the responsibility and evidence boundary at each layer without pretending a static README is a live operations console.
| Layer | Responsibility | Representative systems | Evidence boundary |
|---|---|---|---|
| Observe | Capture protocol traffic, latency, cost, model usage, and business events. | AgentPCAP · TokenGoblin | Raw events and reproducible captures; no inferred health claims. |
| Control | Apply identity, policy, routing, resource, and tenant constraints before execution. | mcpwall · AgentMesh · ModelForge | Versioned policy and explicit inputs; denied actions remain denied. |
| Execute | Run bounded work with idempotency, isolation, retries, and controlled fallback. | JobForge · Reach · local inference | Execution records describe what ran, not what was intended. |
| Prove | Preserve transcripts, provenance, hashes, and replayable evidence. | truthcore · veridag · ReadyLayer | Claims link to inspectable artifacts or remain qualified. |
| Reconcile | Compare technical and financial records against authoritative sources. | Settler · webhook-witness | Provider-correlated settlement is distinct from catalog, checkout, or local state. |
checks the profile structure, generated project metadata, evidence freshness, product pages, local assets, and public links.
These four projects form a complete operating chain and receive the primary attention on this profile. Maturity is deliberately conservative: stable means a versioned release exists; beta means the public system is functional but interfaces may change; research means the architecture is being actively validated.
Public project metadata last verified 2026-09-19 · source manifest · verification policy
| Project | Problem | Public evidence |
|---|---|---|
| AgentPCAP · Go Observe · beta |
Agent failures cross model, tool, MCP, and A2A boundaries that ordinary application logs do not join. | Open .apcap schema, canonical protocol and failure-mode vectors, plus documented CI quality gates. Architecture · Evidence |
| mcpwall · Rust Control · stable |
Tool-capable models need a small, inspectable security boundary before requests reach local MCP servers. | Versioned v1.0.5 release, public firewall tests, release checksums, and a dedicated security workflow. Architecture · Evidence |
| ModelForge · TypeScript Compile · research |
Model deployment choices are usually made through trial, OOM failures, and untraceable sizing assumptions. | Revision-specific compute passports distinguish measured, documented, derived, and predicted evidence; performance CI is public. Architecture · Evidence |
| Settler · TypeScript Reconcile · beta |
Payment, banking, and operational records diverge unless matching and evidence rules are explicit. | Public reconciliation benchmark source and checked-in snapshots, with CI and security-invariant workflows. Architecture · Evidence |
| Artifact | What it demonstrates | Inspect |
|---|---|---|
| AgentPCAP format and vectors | A documented capture container with canonical MCP, A2A, OTLP, retry, incomplete, and error cases. | Format specification · Test vectors |
| mcpwall release and tests | A versioned security boundary with public firewall tests, checksums, and documented limitations. | v1.0.5 release · Firewall tests |
| ModelForge evidence model | Deployment recommendations identify whether inputs are measured, documented, derived, or predicted. | Benchmark schema · Performance CI |
| Settler reconciliation benchmarks | Matching performance is represented by executable benchmark source and checked-in result snapshots. | Benchmark source · Snapshots |
| Profile integrity | The portfolio itself is checked for metadata drift, stale verification, missing assets, malformed product pages, and dead links. | Workflow · Verification script |
The portfolio covers AI infrastructure, enterprise reliability, financial systems, simulation, commerce, and applied tooling. The short list above is curated; the complete map lives here.
AI systems, agent control planes, and inference runtimes
- AgentPCAP
[Go · CLI]— Protocol capture and deterministic replay for AI agents. - AgentMesh
[Go · Distributed]— Identity, policy, routing, reliability, and progressive delivery for A2A and MCP. - ModelForge
[TypeScript · Compute]— Constraint-driven model deployment planning. - mcpwall
[Rust · Security]— Local-first MCP firewall and audit proxy. - veridag
[Rust · Quint]— Formally specified distributed trust DAG. - nlsqlc
[Rust · Compiler]— Multi-tenant natural-language Query IR compiler. - SawyerCore
[Node · Python]— Deterministic edge-AI runtime and simulation engine. - llm-inference-api
[FastAPI]— OpenAI-compatible local inference gateway. - ollama-router
[Python · Daemon]— Multi-lane model routing, health checks, and fallback. - comfyui-api
[Cloudflare · TypeScript]— Headless ComfyUI automation and queue management. - Nautilus
[Docker · Infrastructure]— Containerized operational AI infrastructure. - Keys
[TypeScript]— Auditable mission control for constrained agents. - ControlPlane
[Python · Systems]— Service supervision and operator architecture. - AI-Agent-Portfolio
[Python · Agents]— Agent patterns, tool boundaries, and evaluations. - JupyterNotebooks
[Jupyter · PyTorch]— Quantization, vision, and fine-tuning research.
Governance, verification, and resilience
- continuityos
[Go · OPA]— Sovereign Resilience-as-Code. - FlexibleAccessible
[TypeScript · SaaS]— Continuous accessibility discovery and remediation operations. - ReadyLayer
[TypeScript · CI]— Delivery governance, provenance, and evidence export. - truthcore
[Python · Verification]— Verification kernel, content-addressed cache, and evidence reports. - Reach
[Rust · Runtime]— Deterministic execution and transcript replay. - Requiem
[C++ · Native]— Native execution and operator-console contracts. - JobForge
[PostgreSQL · TypeScript]— Idempotent, RLS-isolated job execution. - MissionLedger
[TypeScript · Policy]— Governed missions, budgets, and proofpacks. - hardonia-compliance-agent
[Rust · Workspace]— Autonomous regulatory-compliance tooling. - hardonia-audit-pack
[Python · Evidence]— Deterministic reconciliation evidence bundles.
FinOps, ledger infrastructure, and commercial engines
- Settler
[TypeScript · TigerBeetle]— Reconciliation intelligence and audit OS. - WhatsForDinner
[React · Stripe]— Consumer AI SaaS with subscriptions, credits, marketplace, and vision workflows. - TokenGoblin
[Go · ClickHouse]— AI token-spend observability and routing guardrails. - apva-framework
[Python · Telemetry]— Reliability-adjusted AI ROI measurement. - finops-autopilot
[Python · FinOps]— Cost anomaly detection and rightsizing policies. - webhook-witness
[Rust · Cryptography]— Signed, tamper-evident webhook ingestion and replay. - commercial-architecture-simulator
[Python · Monte Carlo]— SaaS pricing, churn, and unit-economics simulation. - prompt-ops-hardonia-packs
[Prompt Ops]— Operator packs for GTM, outreach, and verification. - TokPulse
[Turborepo · Remix]— Multi-store creator-commerce operating system. - storefront
[HTML · Edge]— Edge-rendered commerce and digital fulfillment.
Simulation, game runtimes, platforms, and micro-tools
- CEO-G Canada Opportunity Graph
[Python · Graph]— Sovereign infrastructure opportunity modeling. - World26
[Python · Simulation]— Open planetary-systems simulator. - WorldVM
[Rust · WASM]— Sandboxed creator-built gameplay for major game engines. - ReachRadar
[Next.js · Analytics]— Recommendation-algorithm observability. - Zeo
[TypeScript · Edge]— Local-first, signed, composable agent pipelines. - AI Automated Systems
[Astro · Static]— Automation consulting and diagnostic surface. - enterprise-integration-fabric
[Kotlin · Spring]— Governed event-driven integration architecture. - identity-entitlement-broker
[Go · Zero Trust]— Identity brokering and fine-grained entitlements. - api-changelog-radar
[Cloudflare Worker]— Breaking-change detection for vendor APIs. - reliability-platform
[Go]— Circuit breakers and automated disaster recovery. - golden-path-platform
[Terraform · CI/CD]— Compliant internal developer-platform templates. - support-autopilot
[Node.js · CLI]— Autonomous support triage and diagnosis. - ops-autopilot
[Python]— Telemetry-driven reliability proposals through JobForge. - growth-autopilot
[Python]— SEO experiment and content proposals through JobForge. - InboxExorcist
[Python · Gmail]— Reversible inbox decluttering and filter automation. - floyo
[Rust · Telemetry]— Local workflow-opportunity detection. - tfstate-drift-inspector
[Go · Terraform]— Infrastructure drift inspection before apply. - Architecture Playbook
[Documentation]— Public patterns, controls, and evidence boundaries.
The same architecture patterns are packaged as deployable kits, audits, and operator workflows.
| Product | Outcome | Explore |
|---|---|---|
| AI Command Center | Replace operational blind spots with health history, priorities, and revenue-aware triage. | Open |
| SaaS Repo Rescue | Audit auth, billing, webhooks, RLS, security boundaries, and revenue-leaking edge cases. | Open |
| Settler FinOps Engine | Normalize and reconcile payment streams into deterministic evidence packs. | Open |
| TokenGoblin Optimizer | Measure, route, budget, and reduce model inference spend. | Open |
| APVA ROI Benchmark | Calculate reliability-adjusted value before scaling an AI workflow. | Open |
| Local AI Lab Audit | Review GPU utilization, routing, model fit, security, and operating posture. | Open |
| ComfyUI Pro Workflows | Run repeatable, private image-production pipelines on owned compute. | Open |
| Automation Retainer | Add senior architecture and workflow improvement without a full-time hire. | Open |
Explore the complete 28-product catalog
- 8K upscaling and facial restoration
- Architectural and interior visualization
- E-commerce product relighting
- Flux inpainting and outpainting
- Flux portrait studio
- Game asset and PBR texture generation
- Cinematic AI video production
- Voice-to-avatar lip sync
- AI character generator
- AI video storyboard studio
- Consent-based voice cloning
- ComfyUI fashion lookbooks
- ComfyUI product photography
- ComfyUI thumbnail creation
- ComfyUI custom-node starter
| Principle | Working rule |
|---|---|
| Evidence over confidence | If a run cannot be inspected or replayed, it is not production-ready. |
| Local-first by design | Own the compute, data boundary, fallback path, and cost model wherever practical. |
| Determinism at the edges | Keep probabilistic intelligence inside explicit policy, schema, and execution constraints. |
| Boring reliability wins | Idempotency, RLS, state machines, and observable queues beat clever hidden behavior. |
| Revenue is a reconciled event | A dashboard row is not money; provider-correlated settlement evidence is money. |
| Fix the smallest root cause | Isolate the failure, repair it surgically, prove the result, then ship. |
Alongside full-time solutions architecture work at McGraw Hill, I contribute independent, part-time expertise to confidential frontier-AI evaluation and systems initiatives.
The work spans complex technical reasoning, multi-step real-world evaluation design, failure-mode analysis, agent protocol governance, tool-execution boundaries, and structured feedback for enterprise-grade model behavior. Client, model, dataset, and internal research details remain confidential.
If you are working on a serious AI, SaaS, integration, reliability, or revenue system, start with a specific bottleneck, a measurable outcome, and a verifiable path to production.
© Scott Hardie · Hardonia Sovereign Systems · Toronto, Canada