AI Without Fear
Practical local AI tools, grounded retrieval systems, and diffusion workflows built for real hardware.

I build practical local AI systems for people working on real machines, not cloud fantasy hardware.

• Grounded retrieval — source-backed AI answers instead of hallucinated setup steps.
• Local diffusion tools — cleaner image-generation workflows for local hardware.
• Evaluation-first experiments — narrow claims, recorded results, honest limits.

A grounded RAG corpus for local AI workflows.

Covers:

• ComfyUI nodes and API behavior
• Gradio 6 patterns
• model serving
• Python and pip troubleshooting
• evaluation harnesses
• local AI workflow notes

Problem it solves: most AI assistants confidently get ComfyUI node names wrong, recommend outdated Gradio patterns, and hallucinate venv paths. Atlas gives them grounded retrieval material instead.

View AI Without Fear Atlas →

A working research/evaluation lab for testing whether a lightweight QLoRA adapter can learn to read structured Atlas context.

Tests:

• lanes
• cards
• source rules
• compact evidence packs
• selected-card paths
• off-ramp behavior when evidence is missing

Current internal result: the best internal run preserves compact-card behavior, improves targeted retrieval behavior, and records compact selected-card paths using about 5.05x–20x fewer total tokens than raw workspace/RAG-style comparisons in specific lab use cases.

This is not presented as a production package, external benchmark, or universal token-reduction claim.

View Atlas Reader LoRA Lab →

An early public build of a local-first Stable Diffusion workspace and clean-room AUTOMATIC1111-style WebUI rebuild.

Built around:

• typed request and config models
• repo-local runtime folders
• service-routed UI actions
• model management
• txt2img, img2img, and inpaint workflow surfaces
• PNG info, history, settings, enhancement, and segmentation
• early /api/v1 plus A1111-style /sdapi/v1 adapter work

Problem it solves: local diffusion tools are powerful, but many grew quickly into difficult-to-maintain stacks with broad shared state, tightly coupled UI/runtime behavior, and confusing extension paths.

AIWF Studio is not presented as a finished replacement for AUTOMATIC1111, Forge, or ComfyUI.
It is a public foundation for exploring a more maintainable, contributor-friendly local diffusion workspace.

View AIWF Studio →

AI systems should retrieve real project knowledge before answering. Atlas focuses on source-aware retrieval, answer gating, compact evidence packs, and failure behavior when evidence is missing.

AIWF Studio explores what a familiar local diffusion workspace can look like when service boundaries, typed requests, and repo-local runtime behavior are part of the foundation.

The work is built around practical constraints: Windows setups, local models, 16GB/8GB VRAM machines, and workflows normal users can actually run.

The goal is not just to make demos. The goal is to record what works, what fails, and where claims need to stay narrow.

• Local AI — Running and testing models on consumer hardware.
• ComfyUI — Workflows, node behavior, API automation, and datatype boundaries.
• RAG & Retrieval — Corpus design, chunking strategy, source governance, and answer gating.
• Gradio 6 — App building, local UI workflows, and cleaner user-facing tools.
• Model Training — LoRA/QLoRA experiments, evaluation records, and domain-specific adapters.
• Diffusion UX — Making local image tooling easier to understand and maintain.

If Atlas, AIWF Studio, or any of my local AI notes/tools save you time, you can support continued development:

AI Without Fear
Practical local AI for real people, on real hardware.