LumiLabs

Where ideas become infrastructure.

LumiLabs is our internal build environment — active experiments, tools we use daily, and the platforms that power our research.

17+

Active Projects

58+

Engine Modules

Infra Platforms

Languages

Flagship

MIID

Machine Intelligence Infrastructure & Design — our primary research platform and the most ambitious system we build. MIID is a fully local, self-evolving AI development environment that learns from every interaction, reasons across unlimited context, and gets measurably better over time. No cloud. No rate limits. No ceiling.

Architecture

Engine-Based Design

58+ modules each with a defined interface across cognition, memory, AI, IDE, shell, audit, and plugin layers. No monolith, no hidden coupling — every engine independently testable.

Capability

Adaptive Reasoning

MIID doesn’t just respond — it deliberates. Multiple reasoning layers collaborate on complex tasks, each contributing a different depth of analysis. The result is something closer to genuine understanding than pattern matching. True Sentients. True Autonomy. If MIID can’t build it, nothing can.

Delivery

Multi-Surface

Ships as a WinUI 3 native IDE, Blazor web IDE, CLI, system tray, and REST API — all driven by the same core engine layer, compiled with native AOT for peak performance.

Projects

Active Lab Inventory

Everything currently running or in active development.

Inference Research

PILOT

A management and interface layer for locally-hosted Ollama models. Routes requests, manages model lifecycle, and abstracts provider differences across deployments.

Local LLM Ollama .NET

Orchestration In Dev

Maestro

Agent orchestration platform for coordinating multi-step AI workflows. Handles task delegation, dependency resolution, and result aggregation across agent boundaries.

Agents Workflow Orchestration

Inference Research

SiliconFlow Toolset

A growing set of integration tools for working with the SiliconFlow AI provider. Covers API interaction, model routing, and workflow automation built on top of their platform.

Integration SiliconFlow API Tooling

Data In Dev

LumiSQL

Enterprise SQL Server monitoring and observability platform with ML-powered anomaly detection, real-time diagnostics, and autonomous healing — zero external infrastructure.

SQL Server ML.NET .NET 10

Learn more →

Storage In Dev

SprawlFS

Distributed fault-tolerant file storage built to survive node failures without flinching. Shards data across independent nodes so files stay recoverable even when infrastructure goes dark.

C++20 Distributed Fault-Tolerant

Learn more →

Automation Active

MagicMacro

Keyboard and workflow macro automation designed for developer contexts. Reduces repetitive actions across IDEs, terminals, and research tooling with minimal config.

Automation Productivity Input

Messaging Research

LumiMIM

Message-in-motion infrastructure for inter-service communication. Designed for the low-latency, high-reliability demands of our multi-engine platforms.

Messaging IPC Transport

Observability Active

ClaudeWatchdog

Session monitoring and observability tool for Claude AI interactions. Captures, stores, and surfaces patterns from agent sessions over time.

Monitoring Claude Sessions

Security Research

InterceptAlloy

Network and data interception framework for security research and traffic analysis. Used to understand what flows through systems before designing mitigations.

Security Interception Analysis

Interface Research

OdigoII

Adaptive navigation and guidance system that surfaces relevant context based on current user state, reducing the cognitive cost of context switching across complex toolsets.

Navigation Context Interface

AI / Inference In Dev

.NetTransformer

Complete transformer training stack built from first principles in pure .NET — combined with S.Q.E.E.Z semantic inference, four desktops match what used to require a datacenter.

.NET 10 C# Native Distributed Swap-File

Learn more →

Experiment Active

HyperVibing

Early-stage experiment in rapid AI-assisted prototyping workflows. Exploring how to compress the gap between idea and working prototype using scaffolded generation.

Prototyping Scaffolding AI-Assisted

Data In Dev

DotNetDataSetScraper

Targeted dataset collection tooling for .NET ecosystem resources. Used to build training and evaluation corpora for MIID’s CoreTeachingEngine.

Data .NET Training

Blockchain Research

LumiCoin

A production-grade blockchain platform in .NET 8. Proof-of-Work consensus, UTXO state manager with O(1) balance lookups, BIP-39/BIP-32 HD wallet with AES-256 encryption, ring signatures, stealth addresses, and 5-level coin mixing for privacy-by-default. Built-in multi-arbiter escrow with reputation scoring and auto-refund timeouts. TCP-based P2P network with gossip protocol, block sync, peer discovery, and DDoS mitigation. AI-powered real-time transaction threat detection with federated learning. Cross-platform Avalonia desktop client and a JWT-authenticated REST + SignalR API. 381 passing tests.

Blockchain Privacy AI Security .NET 8 Escrow P2P

Inference Research

S.Q.E.E.Z

Stream Quantization Engine — reduces AI model memory 75–80% and first-inference latency 60–75% through semantic chunking, predictive prefetching, and an autonomous self-healing AI consultant.

Quantization Inference Python

Memory Research

Lumi-NEH-Sense

Neural Enhancement Hub — a lightweight, plugin-based persistent memory system for AI applications. Ultra-simple core (~5 dependencies), 30+ planned plugins covering storage, security, monitoring, and LLM integrations.

Neural Memory Plugins .NET 8

AI Coding Active

Ogosho

“Grand Master.” A local AI coding assistant for Windows — modern console, retro TUI, and a native Avalonia IDE, all driven by one shared agent runtime, and the seed of a framework destined to carry every Lumiotic product into AI readiness.

Ollama Avalonia .NET 10 Local LLM

Learn more →

Process

How We Build

LumiLabs operates in focused build cycles. Each cycle targets a specific capability gap identified through research or operational need.

We build for ourselves first — solving real problems with real constraints — and release selectively when the work might benefit others.

Focused cycles with clear scope before starting
Internal dogfooding before any external release
Everything instrumented and observable from day one
Documentation ships alongside the code

17DMV Collectors

8Projects / Layers

<1%CPU Overhead

What it is

LumiSQL is a pure .NET SQL Server monitoring and observability platform — the kind enterprises pay six figures a year for, rebuilt from scratch with no compromises on stack complexity. It runs entirely on a single Windows box with no external message queues, no Python, and no cloud dependency. Install it, point it at a SQL Server instance, and it starts learning.

Where commercial tools bolt on AI as a feature, LumiSQL is designed intelligence-first: ML.NET-powered anomaly detection and 24-hour forecasts are built into the collection pipeline, not layered on top afterward. Alerts stop being static thresholds and start being learned baselines that adapt to your workload’s real behavior over time.

Distributed deployments are handled by a lightweight Windows Service agent that collects locally and posts to a central server — no Kafka, no RabbitMQ, no infrastructure to babysit. The agent buffers to disk if the server is unreachable and drains automatically when it reconnects.

Architecture

click to enlarge

Capabilities

17 DMV Collectors CPU, memory, wait stats, blocking chains, deadlocks, Query Store, AG replication lag, TempDB, disk I/O, SQL Agent jobs — captured every collection cycle with <1% server overhead.

ML.NET Anomaly Detection SSA spike and change-point detection on every metric stream. No Python, no external AI service. Interpretable, deterministic statistical models that ship inside the binary.

24h / 7d Forecasting Capacity forecasts with confidence intervals derived from learned hourly, daily, and weekly workload patterns. Alerts can fire on forecast breach — not just current threshold.

Real-time Dashboard React 18 SPA with 17 diagnostic pages. SignalR pushes metric updates in under 3 seconds from collection. Subscriptions are instance-scoped to prevent fan-out storms at scale.

Autonomous Healing Kill blocking sessions, shrink logs, resolve deadlocks — with admin-gated approval before execution. Full history of applied actions tied to the metric events that triggered them.

Zero Infrastructure SQLite for config, DuckDB for time-series. No external databases, no message queues. Distributed agents use HTTP + disk buffering. Single self-contained win-x64 executable.

Stack

.NET 10 / C# 13 ASP.NET Core React 18 TypeScript SignalR MediatR / CQRS ML.NET 4.0 DuckDB SQLite + EF Core Dapper Polly Serilog Windows Service xUnit · FluentAssertions Testcontainers.MsSql

∞Scale Horizontally

Any NodeCan Recover the Whole

0Backups Required

What it is

SprawlFS is a distributed file storage system engineered around one principle: files should survive infrastructure failure without you having to think about it. Data is fragmented and encoded redundantly across independent storage nodes, so as long as a minimum threshold of nodes remain reachable, every file in the system stays fully recoverable — even while nodes are actively offline.

Most storage systems treat node failure as an emergency. SprawlFS treats it as a normal operating condition. The system transitions gracefully across three observable states — healthy, degraded, and critical — and the built-in web dashboard makes that state visible in real time: which nodes are up, which files are at risk, and exactly how much headroom you have before recovery becomes impossible.

Files are immutable on write. Every upload creates a new version, with older versions retained until explicitly purged. That means accidental overwrites are recoverable, and you get a full timeline of every file in the system — not just the latest copy.

Architecture

click to enlarge

Capabilities

Survive Node Failures Files remain fully recoverable as long as the minimum threshold of storage nodes stays reachable. Node loss is a normal operating state, not a disaster.

Three Observable Health States Healthy, Degraded, and Critical — each with precise, per-file visibility. You always know exactly how much fault tolerance headroom remains before recovery is at risk.

Automatic Version History Every write is immutable. Previous file versions are retained automatically, giving you a full timeline you can recover from on demand — not just the latest copy.

Streaming Uploads Large files stream directly to storage without full in-memory buffering. Memory footprint stays flat regardless of file size.

Live Health Dashboard Web UI shows node status, per-file health indicators, and version history in real time. Toggle nodes offline to verify survivability before you need it to matter.

Self-Contained Binary Single native executable. No runtime, no external dependencies, no infrastructure to provision. Runs on bare metal or in a VM — wherever you need storage to survive.

Stack

C++20 Windows Native WinSock HTTP Server SHA-256 Content Hashing Erasure Coding Append-only Metadata Log Immutable Versioned Storage Vanilla JS Dashboard REST API

75-80%RAM Reduction

10,000×Cached Speedup

0External ML Deps

What it is

.NetTransformer is the world’s first complete transformer training stack built from first principles in pure .NET — no Python, no PyTorch, no external ML framework dependency. Every layer, from the tensor substrate and automatic differentiation engine through the training loop, optimizer, and distributed coordination protocol, is written in C# and fully owned. If the capability doesn’t exist in the ecosystem, it gets built.

Combined with S.Q.E.E.Z’s semantic inference engine, it becomes something the industry hasn’t seen: a complete AI platform that trains domain-specific models and serves them at scale without GPU clusters or cloud infrastructure. S.Q.E.E.Z reduces active memory requirements by 75–80% and delivers sub-second inference on hardware most organizations already own — with accuracy that improves automatically as the system learns from use.

Together, these two engines answer the question the AI industry has been avoiding: does intelligence at scale require $50,000/month in GPU infrastructure? Four desktops running this stack match the practical throughput of configurations costing 100× more — not by cutting corners, but by rethinking where computation actually needs to happen. No giganormous datacenters. Just a few machines that think at datacenter scale.

Architecture

click to enlarge

Capabilities

Training on Consumer Hardware Empirically faster training than conventional in-memory approaches — measured, not claimed. No GPU cluster required. Scales across multiple machines using a round-based distributed protocol that prioritizes local compute over network chatter.

75–80% Memory Reduction at Inference Active model footprint shrinks dramatically through selective weight loading. Only the components relevant to each specific input are in memory at any time — determined before inference begins, not reactively during it.

Pattern-Adaptive Inference 85–95% cache hit rate on real-world traffic. Response time for repeated patterns measured in milliseconds. The system improves automatically through use — no retraining, no redeployment, no manual tuning required.

Distributed Training Without Infrastructure Round-based protocol coordinates training across multiple machines with integrity validation at every merge. Deterministic, reproducible, and resilient to node failure. No message queues, no cloud coordination, no ops overhead.

Complete .NET Transformer Stack Generic tensor arithmetic, reverse-mode automatic differentiation, multi-head attention, full transformer architectures, and production-grade optimizers — all in C#. Zero Python. Zero PyTorch. Zero external ML framework.

Desktop Datacenter Four consumer machines running this combined stack produce AI throughput that previously required dedicated GPU infrastructure costing 100× more. Not a benchmark optimization — a structural property of how the engines were designed from first principles.

Stack

.NET 10 / C# 13 Generic Tensor<T> Reverse-Mode Autograd Multi-Head Attention ILGPU / CUDA Semantic Chunking Adaptive Pattern Cache Optimized Tensor Store Distributed Merge Protocol Safetensors Direct Access

4Surfaces · One Runtime

0Cloud Dependencies

FrameworkFoundation for the Lineup

What it is

Ogosho — Japanese for Grand Master — is a local AI coding assistant that runs entirely on your Windows machine. It pairs an Ollama-driven reasoning core with a deliberate, deeply-integrated tool surface: shell execution, filesystem operations, regex content search, web fetches, and live inspection of the windows currently open on your desktop. The result is an agent that actually does work — not one that merely talks about it.

The same agent runtime sits behind every Ogosho experience: a modern console with rich formatting and clipboard handling, a full retro Terminal.Gui TUI for keyboard-heavy workflows, and Ogosho.Studio — a native Avalonia desktop IDE with multi-file document tabs, clickable file mentions in chat, model and host selectors, and a live session feed. Switch surfaces mid-task and the work continues exactly where you left it. One identity. One history. One set of tools. Four ways to talk to it.

Ogosho is also the first incarnation of a broader framework Lumiotic is quietly building — a foundation designed to carry every product in the lab into AI readiness as the lineup matures.

Architecture

click to enlarge

Capabilities

Local LLM Reasoning Powered by Ollama running on your machine. Your code, your prompts, and your session history never leave the box. Swap models per session with one command — llama3, qwen, whatever your hardware supports.

One Runtime · Four Surfaces The console, the retro TUI, Studio, and the REST server all run the same agent runtime. Same tool catalog, same command vocabulary, same conversational behavior — verified by parity, not coordinated by convention.

Native Avalonia Studio Multi-file document tabs, clickable file mentions, model and host selectors, a live event feed, and a runtime panel showing the platform’s effective listen URL, transport mode, and auth posture. Code-editor evolution in active development.

AI-Readiness Framework Ogosho is the first surface of a broader framework Lumiotic is building. The runtime, tool catalog, and parity model proven here will, over time, become the foundation that prepares every product in the lab — MIID and beyond — to converse with and act on behalf of their users.

Tool-Using Agent Shell execution, file read/write/edit, directory listing, regex content search, web queries, URL fetch, and visible-window inspection — all callable by the model with per-turn user confirmation by default. Override with full-throttle when you trust the loop.

Persistent Sessions Conversation history is preserved across launches. Save and load named sessions, resume work seamlessly, and switch surfaces without losing context.

Stack

.NET 10 / C# 13 Avalonia UI Terminal.Gui Ollama System.CommandLine Spectre.Console ASP.NET Core Minimal API Single-File Publish Windows Native Lumiotic AI Framework