Most of it is pulling throughput from the GPUs you have and routing each request to the model that should handle it.
We build sub-10ms semantic search for conversational AI, shaving at least 200-300ms off voice AI stacks.
What surprised me: how many teams reached for Moss as their retrieval layer, and how much could ship as-is.
One covers GPU limits, VRAM and quantization. The other, Python threading and caching that cut API costs up to 90%.
At Kelley I probed Hugging Face LLMs for security holes. I also led lectures for a 70-student AI course.
It runs RAG pipelines on a fine-tuned LLaMA 3.2 11B, with an agent watching for prompt injections.
I streamed Whisper STT and TTS over WebSocket, and helped fine-tune Llama 3.1 to rely less on ChatGPT's API.
At Quidich I forecast ball release from player coordinates. At IDeaS, automation saved 500+ hours of manual work.
I build low-latency retrieval for voice apps and high-throughput model serving systems.
Most of it is pulling throughput from the GPUs you have and routing each request to the model that should handle it.
We build sub-10ms semantic search for conversational AI, shaving at least 200-300ms off voice AI stacks.
What surprised me: how many teams reached for Moss as their retrieval layer, and how much could ship as-is.
One covers GPU limits, VRAM and quantization. The other, Python threading and caching that cut API costs up to 90%.
At Kelley I probed Hugging Face LLMs for security holes. I also led lectures for a 70-student AI course.
It runs RAG pipelines on a fine-tuned LLaMA 3.2 11B, with an agent watching for prompt injections.
I streamed Whisper STT and TTS over WebSocket, and helped fine-tune Llama 3.1 to rely less on ChatGPT's API.
At Quidich I forecast ball release from player coordinates. At IDeaS, automation saved 500+ hours of manual work.
I build low-latency retrieval for voice apps and high-throughput model serving systems.
Most of it is pulling throughput from the GPUs you have and routing each request to the model that should handle it.
We build sub-10ms semantic search for conversational AI, shaving at least 200-300ms off voice AI stacks.
What surprised me: how many teams reached for Moss as their retrieval layer, and how much could ship as-is.
One covers GPU limits, VRAM and quantization. The other, Python threading and caching that cut API costs up to 90%.
At Kelley I probed Hugging Face LLMs for security holes. I also led lectures for a 70-student AI course.
It runs RAG pipelines on a fine-tuned LLaMA 3.2 11B, with an agent watching for prompt injections.
I streamed Whisper STT and TTS over WebSocket, and helped fine-tune Llama 3.1 to rely less on ChatGPT's API.
At Quidich I forecast ball release from player coordinates. At IDeaS, automation saved 500+ hours of manual work.
By using this service, you agree to the Terms of Service and Privacy Policy.