Shane Logsdon
@shane.logsdon.io
Founder @ LeadSurface · Surfacing in-market buyers from the forums where they’re already talking · Developer Advocacy @ Global Payments · buying signals, intent data
Beware if you try it. Skills Pi loads from directories you haven't listed leave your config and don't enter search. You have to tell it where to look.
MIT: github.com/slogsdon/pi-...
MIT: github.com/slogsdon/pi-...
GitHub - slogsdon/pi-deferred-skill-catalog: Pi extension for deferred skill loading: keep skill descriptions out of the system prompt, search them on demand, and expand /skill: commands.
Pi extension for deferred skill loading: keep skill descriptions out of the system prompt, search them on demand, and expand /skill: commands. - slogsdon/pi-deferred-skill-catalog
github.com
September 21, 2026 at 1:05 AM
Beware if you try it. Skills Pi loads from directories you haven't listed leave your config and don't enter search. You have to tell it where to look.
MIT: github.com/slogsdon/pi-...
MIT: github.com/slogsdon/pi-...
Next - blind human review, retrieval tests, an enforced-JSON variant, more repetitions, and the same weights on real GPU/NPU hardware.
Methodology, code and raw results: github.com/slogsdon/local-ai-for-smb
If you run local inference or support SMB IT: which workload would you test first?
Methodology, code and raw results: github.com/slogsdon/local-ai-for-smb
If you run local inference or support SMB IT: which workload would you test first?
GitHub - slogsdon/local-ai-for-smb: Applied research into when local AI is a practical, private, and economically compelling option for small and midsize businesses.
Applied research into when local AI is a practical, private, and economically compelling option for small and midsize businesses. - slogsdon/local-ai-for-smb
github.com
September 16, 2026 at 11:56 PM
Next - blind human review, retrieval tests, an enforced-JSON variant, more repetitions, and the same weights on real GPU/NPU hardware.
Methodology, code and raw results: github.com/slogsdon/local-ai-for-smb
If you run local inference or support SMB IT: which workload would you test first?
Methodology, code and raw results: github.com/slogsdon/local-ai-for-smb
If you run local inference or support SMB IT: which workload would you test first?
I can't claim any provider or hardware ranking, a latency comparison (different models, network overhead), or cost break-even yet. Human review is still pending.
September 16, 2026 at 11:56 PM
I can't claim any provider or hardware ranking, a latency comparison (different models, network overhead), or cost break-even yet. Human review is still pending.
The two failed for different reasons. The small local model broke format and mixed up Standard and Premium coverage. The larger hosted model kept format but over-inferred.
That's a hypothesis for hybrid routing, not a finding yet.
That's a hypothesis for hybrid routing, not a finding yet.
September 16, 2026 at 11:56 PM
The two failed for different reasons. The small local model broke format and mixed up Standard and Premium coverage. The larger hosted model kept format but over-inferred.
That's a hypothesis for hybrid routing, not a finding yet.
That's a hypothesis for hybrid routing, not a finding yet.
The hosted model, DeepSeek v4 Flash on NIM, returned valid JSON on every successful call.
It still said contractors get no meal allowance. The source only says contractors aren't covered by that policy. It cited the source anyway.
It still said contractors get no meal allowance. The source only says contractors aren't covered by that policy. It cited the source anyway.
September 16, 2026 at 11:56 PM
The hosted model, DeepSeek v4 Flash on NIM, returned valid JSON on every successful call.
It still said contractors get no meal allowance. The source only says contractors aren't covered by that policy. It cited the source anyway.
It still said contractors get no meal allowance. The source only says contractors aren't covered by that policy. It cited the source anyway.
First lesson. An HTTP 200 isn't task success.
Gemma 3 4B on an M4 Mac returned 0 of 16 strictly valid JSON responses on both local runtimes. Every answer came wrapped in Markdown fences. I didn't strip them or fix the prompt afterward.
Gemma 3 4B on an M4 Mac returned 0 of 16 strictly valid JSON responses on both local runtimes. Every answer came wrapped in Markdown fences. I didn't strip them or fix the prompt afterward.
September 16, 2026 at 11:56 PM
First lesson. An HTTP 200 isn't task success.
Gemma 3 4B on an M4 Mac returned 0 of 16 strictly valid JSON responses on both local runtimes. Every answer came wrapped in Markdown fences. I didn't strip them or fix the prompt afterward.
Gemma 3 4B on an M4 Mac returned 0 of 16 strictly valid JSON responses on both local runtimes. Every answer came wrapped in Markdown fences. I didn't strip them or fix the prompt afterward.
Phase 0 is done. A reproducible harness running one synthetic SMB knowledge-assistant workload, unchanged, through Ollama, Lemonade and hosted NVIDIA NIM.
No per-provider prompt tuning. Failures stay in the raw results.
No per-provider prompt tuning. Failures stay in the raw results.
September 16, 2026 at 11:56 PM
Phase 0 is done. A reproducible harness running one synthetic SMB knowledge-assistant workload, unchanged, through Ollama, Lemonade and hosted NVIDIA NIM.
No per-provider prompt tuning. Failures stay in the raw results.
No per-provider prompt tuning. Failures stay in the raw results.
Interesting. Will have to take a look
August 1, 2026 at 5:37 PM
Interesting. Will have to take a look
Want a few people to kick the tires before I share it wider. DM me. Story here: shane.logsdon.io/articles/str...
Building on the Margins
A founder with a day job and family builds LeadSurface via a Loop & Gate workflow, keeping product strategy human, handing disciplined execution to a pipeline.
shane.logsdon.io
July 14, 2026 at 8:57 PM
Want a few people to kick the tires before I share it wider. DM me. Story here: shane.logsdon.io/articles/str...