AI infrastructure and agentic systems engineer building local-first, auditable software for real operations.

I design and operate InnerOS: a governed multi-agent operating layer that connects software delivery, infrastructure, business workflows, persistent memory, and human approval. The objective is practical—return human attention by automating coordination, verification, recovery, and repetitive operational work.

• Agentic infrastructure: durable multi-agent coordination, MCP/A2A integrations, scoped tools, task ownership, locks, evidence, and recovery.
• Local AI systems: private inference across NVIDIA and AMD/ROCm nodes, with capability-aware routing and cloud burst only when justified.
• Operational products: workforce, payroll, service operations, security, automation, and AI-assisted decision systems grounded in real company workflows.
• Governed execution: deterministic business rules, least privilege, human approval gates, bounded spending, and verifiable outcomes.

InnerOS is not another chatbot. It is an operating layer for turning signals into safe, reviewable action:

signal → relevance → memory → decision → delegation
       → execution → verification → recovery → evidence → learning

The system is designed around a simple truth: agents, services, models, and humans all fail. Reliable autonomy therefore requires persistent state, explicit authority, failure detection, bounded repair, and evidence before completion.

Human / business / infrastructure signals
                  │
                  ▼
        InnerOS coordination plane
    identity • memory • policy • routing
                  │
       ┌──────────┼──────────┐
       ▼          ▼          ▼
 local models   cloud AI   specialist agents
       │          │          │
       └──────────┼──────────┘
                  ▼
 scoped tools • Git worktrees • business systems
                  │
                  ▼
 verification • recovery • audit evidence

_{InnerOS architecture. Perception, reasoning and action are connected through governed execution, evidence and local-first compute.}

InnerOS is developed as one system with several layers. Public repositories fall into one of these roles:

The relationship is deliberate:

real operational problem
        ↓
InnerOS core capability
        ↓
bounded experiment / hackathon validation
        ↓
measured evidence
        ↓
reusable capability
        ↓
maintained product or platform layer

Hackathons are therefore part of the R&D process, not separate product identities. Submission repositories preserve the historical evidence; reusable engineering moves forward into maintained InnerOS components.

_{R&D as a system. Hackathons and bounded experiments validate technology; only useful, evidenced capabilities are promoted into maintained InnerOS products or core infrastructure.}

These six repositories best explain the current InnerOS system from platform to real-world operation:

Together they show the full stack:

coordination
    ↓
memory + reasoning
    ↓
perception + interfaces
    ↓
governed execution
    ↓
real-world operations
    ↓
evidence + safety

• Workforce — attendance, incidents, reporting and deterministic pre-payroll automation.
• VoiceOps — governed voice interaction for operational actions, approval and evidence.
• FieldOps — human-approved workflows for real-world tasks and physical operations.
• Ambient Guardian — Alexa+/MCP orchestration for contextual, verified smart-environment actions.

• Forensic Replay — content-addressed evidence, deterministic replay and counterfactual analysis.
• DMX Engine — bounded physical-control execution for lighting and stage automation.
• MCP / A2A infrastructure — agent-to-tool and agent-to-agent control surfaces across the ecosystem.

AMD/ROCm, HyperLoom, AWS, AssemblyAI, AI Infra Summit, Executable World and other hackathons are treated as bounded R&D environments. They validate technologies and architectural hypotheses; successful capabilities are promoted into maintained InnerOS products or core infrastructure.

1. Local-first, not local-only. Use private infrastructure by default; use cloud capacity when it creates measurable value.
2. Evidence over confident claims. A task is not complete without tests, runtime evidence, or an explicit truth boundary.
3. Authority is part of the architecture. High-impact actions require scoped permissions, approvals, and auditability.
4. Failure is a normal state. Detect stalls, prevent collisions, recover safely, and preserve what happened.
5. Human attention is the metric. Optimize for time returned, errors avoided, decisions prepared, and work actually closed.
6. Freeze → Extract → Integrate. Preserve hackathon submissions as historical artifacts; move reusable capabilities into maintained successors instead of rewriting the evidence.

• MCP and A2A control planes for heterogeneous agent fleets
• local model routing with AMD ROCm/vLLM and NVIDIA/Ollama infrastructure
• durable coordination, forensic replay, and recovery automation
• secure Git-based development through isolated branches and worktrees
• deterministic operational workflows backed by MongoDB, Qdrant, and Notion
• cost-aware cloud burst and short-lived infrastructure
• workforce, service operations, payroll, physical security, and automation

This account contains several kinds of repositories:

• Canonical platform and products — maintained InnerOS and business capabilities.
• Hackathon submissions — frozen snapshots that preserve what was submitted.
• Successor projects — continued development extracted from those snapshots.
• Experimental probes — bounded repositories used to prove or reject one technical hypothesis.
• Public evidence and identity — profile, engineering journal, demos, and documentation.

That distinction is intentional. Hackathon repositories are not silently rewritten after submission; reusable engineering is extracted into successor repositories with clearer ownership and lifecycle.

• prove end-to-end autonomy through real operational workflows rather than agent demos;
• consolidate the InnerOS architecture and repository lineage into a recoverable system map;
• turn Workforce and Service Operations into production-quality products;
• strengthen security, observability, forensic replay, and self-healing behavior;
• measure human time returned and operational work closed.

• InnerChispa: innerchispa.us
• GitHub: github.com/Rafa-Innerchispa

Building AI systems that notice what matters, do what they safely can, verify the result, and return the human only when human judgment is genuinely required.