I build agents and fine-tune models. Mostly in the Hermes ecosystem: the model itself, the provider adapters it runs on, the context engine behind it.
My gpt-oss-20b tool-use finetune ships in four runtimes with 9k+ downloads. Two of my fixes are merged into hermes-agent, one into hermes-lcm.
role Data & ML Engineer
focus agent infra · fine-tuning · adapter correctness · context
shipped gpt-oss-20b tool-use finetune — MLX · GGUF · 4-bit · 16-bit
adoption 9k+ Hugging Face downloads
oss merged fixes in repos totaling 205k+ stars
based Miami · B.S. Data Science & AI, FIU
gpt-oss-20b, tuned for Hermes-style function calling. Same weights, four runtimes:
| model | format | runs on |
|---|---|---|
finetune_mlx |
MLX | Apple Silicon |
finetune_gguf |
GGUF | llama.cpp · Ollama · LM Studio |
finetune_4bit |
4-bit | Colab · low-VRAM |
finetune_16bit |
16-bit | vLLM · full precision |
the collection → · repo notes →
Five fixes merged across hermes-agent and hermes-lcm (205k+ stars combined). These bugs only surface in real agent runs: signed-reasoning replay, tool-use stripping, context assembly.
-
hermes-agent— Anthropic extended-thinking crash loop
Authored the fix for crashes when orphan tool-use stripping invalidated signed reasoning blocks. Teknium cherry-picked it tomainwith authorship preserved.64628ea· closes #35847 -
hermes-agent— WhatsApp LID alias resolution
Authored the session-path fix so modernplatforms/whatsapp/sessioninstalls stop silently dropping allowlisted senders. Accepted with authorship preserved.263ffec· closes #36664 -
hermes-agent— interleaved-thinking signature crash loop
Filed and root-caused; maintainers confirmed the order-preserving channel onmainsolved it without demoting reasoning. -
hermes-agent— leading-assistant transcript backstop
Maintains an adapter-level backstop PR for post-compaction transcripts, complementing the built-in compressor fix for #52160 / #52167. -
hermes-lcm— DAG context role invariant
Merged fix so DAG context summaries cannot become provider-visible leading assistant messages.
More detail: oss-contributions.
gpt-oss-20b-hermes-tool-finetune— end-to-end tool-use fine-tuning notes, export matrix, local inference packaginghermes-Notion-Worker-sync— Discord ↔ Notion worker sync on@notionhq/workersamazon-image-quality-bsr-analysis— applied DS: image-quality metrics, pricing/review enrichment, BSR modelingfesalfayed.com— my site, a tmux session in the browser
fix the root, not the symptom · local default, cloud on demand
agents type, humans steer · one project, one repo, one job
hi@fesalfayed.com · huggingface.co/fesalfayed · fesalfayed.com
Data & ML Engineer · Miami · B.S. Data Science & AI, FIU