PORTFOLIO / AI AND AUTOMATION

How I use local AI and save my budget for complex work

For routine work I can verify, I run a language model (LLM) on my PC. The model is Qwen3 and I run it with Ollama. That gives me more room to use advanced models when I need to analyze a situation, make a decision or solve a difficult problem.

A meeting has just ended. I need to know who committed to doing what, and by when. If the transcript says it clearly, I have the model on my PC prepare the notes. If two meetings contradict each other and a decision depends on the answer, I bring the relevant excerpts to a more advanced model. For a team, this keeps routine follow-ups moving without using up capacity for the work that needs judgment.

How I choose the model

Define the taskIs the output specific and easy to check?
Run it on the PCExtract, classify or rewrite using the supplied sources.
VerifyI check the important excerpts and verify the numbers.
Use an advanced modelWhen sources conflict, the context is too long or the decision has consequences.

I choose the model for each task and decide when to switch to an advanced one.

The Ollama and Qwen3 logos side by side on a computer screen with the words Local AI.

What this looks like in my day

Configured workflow

Summarize a meeting

From this transcript, separate decisions, proposals and commitments. For each commitment, give the owner and due date if stated.

Screenpipe turns the audio into text. During setup, I got Whisper (Large V3 Turbo) ready to transcribe on my PC. My current setup uses Screenpipe’s cloud service for that step. Then Qwen3 runs on my PC through Ollama to summarize the transcript, preserve my notes and flag missing details. I check the commitments before following up.

Example of my routing rule

Prepare a marketing update

Group these observations by campaign and draft a comment from the differences already calculated.

A script calculates the changes in spending and results. The local model can organize the observations and draft a first summary. If the question becomes “Why are results down and where should the budget go?”, I reserve that analysis for an advanced model, with the source data.

Example of my routing rule

Prepare a small Power Apps fix

Adapt this Power Fx formula to display a message when the field is empty. Give me three cases to test.

I can ask the local model for a short proposal and test cases. I check the behaviour in the app. A change that affects several screens, navigation or access calls for a deeper analysis.

What I am trying to save

A request handled on my PC does not call a hosted model. With a metered API, that can avoid a processing charge. With a fixed subscription, the price stays the same. I preserve usage capacity so I can choose an advanced model more often when the work warrants it.

I do not have measured savings yet. To measure them, I would count the tasks completed locally, the ones redone in the cloud, the time spent checking outputs and the electricity used. If I have to redo the same task several times, the gain can disappear. That is why I start with short, repeatable tasks that are easy to check.

Recommended systems and hardware

These are starting points for text models and one person at a time. Start with the machine you already have. Requirements vary with the model, its compressed version and document length.

I started with Windows 11, 32 GB of RAM, a 10 GB RTX 3080, Ollama and a 4-bit Qwen3 8B. I verified that the model loaded entirely onto the GPU with my test settings.

Try it on an existing PC

System and memory: Windows or Linux, 16 GB of RAM and 20–30 GB of free SSD space.
Graphics and models: A compatible card with 4–6 GB of VRAM for small 2–4 billion parameter models (2–4B) and short inputs. Without a dedicated GPU, try a small model on the CPU and check the speed.

Use it regularly on a PC

System and memory: Windows or Linux, 32 GB of RAM. Allow 50 GB of free SSD space if you keep several models.
Graphics and models: Consider a compatible NVIDIA card with 12–16 GB of VRAM for 4-bit 7–8B models and trying a 14B model with modest context. For AMD, check the specific card and supported drivers.

Use an Apple Silicon Mac

System and memory: macOS, 16 GB of unified memory to start. Consider 24–32 GB for more headroom with other apps open.
Graphics and models: The processor and graphics hardware share memory. Start with a small model, then try a 4-bit 7–8B model if the machine can handle it.

Which software?

Ollama on Windows, macOS or Linux to run the model and connect it to a workflow. LM Studio is an option if you prefer a graphical interface for testing models. Check exact hardware compatibility before buying.

When choosing a graphics card, allow for the memory needed to process your text. The 4-bit Qwen3 8B download is about 5.2 GB. Longer inputs and other software use memory too. If you are aiming for models around 30B, look at cards with 24 GB of VRAM or more and test your intended model and context before investing.

Technical references

What I keep under control

The local model works from what I give it. A summary can miss a nuance, and an advanced model can make mistakes too. I check the passages that matter, leave arithmetic to scripts and verify the result in the source application. If a document is too long, I split it for simple extraction or change models if the answer depends on the full record. In my meeting workflow, the summary runs on my PC. If I send an excerpt to a hosted model, I follow the project’s data access rules.

Start with one recurring task

If you want to try this with your team, get in touch. We can start with a meeting recap or a recurring export, test the results on your hardware, and decide which requests to keep local.

Discuss a first use case