From US$100 to under US$8 per working day.

Daily AI spending had reached US$100. As the team used agents for more work, usage limits became an operational constraint. The choices included waiting for a reset, buying additional capacity or managing growing consumption-based costs.

The work needed a way to allocate tasks that supported continued delivery and made higher-capability models available where their contribution mattered most.

I kept planning, difficult decisions and output supervision with the stronger models, and built a shared hosted-model path for suitable execution tasks.

I assessed options for reasoning, output quality on relevant tasks, operating requirements, image support and context length. I configured LocalAI and llama.cpp on two NVIDIA L4 GPUs, with image support and a context window of up to 256K tokens.

The purpose was to improve operating economics while retaining task selection, review and human responsibility for the work.

I configured integration with Codex and Claude Code, putting the hosted model into the team’s execution workflow. As part of Tahliq’s wider AI setup, I also developed custom skills and MCP integrations connecting agents to their tools, and trained the team to use them.

The work required decisions about the task, the capability it needed, the model and its environment, and the tools and context available to the agent. I continued to direct execution, assess the output and address obstacles as they appeared.

Keeping Codex and LocalAI working together

Codex exposed gaps in LocalAI’s WebSocket response handling: warm-up requests needed to complete without generating output, and continuing a conversation needed its full history. Response storage also needed to keep each caller’s data isolated.

I traced the problem, directed AI agents through the repair and reviewed the implementation and checks. The contribution covered warm-up handling, complete conversation continuation, response ownership and failure sequencing.

Regression tests exercised connection isolation, successive conversation turns, concurrent requests and failures during streaming. The fix, its tests and the accompanying usage documentation were accepted and merged into LocalAI.

LocalAI compatibility contribution — #11608

Checking the guidance against actual operation

The documentation said speculative decoding was disabled when image processing was active. I checked that claim against the backend and tested a configuration combining vision input with Multi-Token Prediction (MTP).

Successful image requests and runtime logs showed both capabilities operating together. I submitted corrected guidance, the configuration example and the checks operators could use to verify their own setup. This contribution was also accepted and merged.

LocalAI documentation correction — #11700

Selected contributions to AI tools

Daily operating costs fell from US$100 to under US$8, a reduction of more than 92%, while serving 4 people through the shared environment. Stronger models remained in use for planning, supervision and harder tasks.

The project brings together several levels of my AI expertise: model selection and hosting, agent integration, task direction, output review and following a problem into the original tool when the solution requires it.

In marketing, that capability expands delivery across research, analysis, production and automation. Elkhabar Bewodooh shows another application, through the filming and publishing tools I built around its content strategy.

Explore the media production application

Review my agency experience · Discuss your team’s needs