christians site

ai/ml/ds

i have done a bit of machine learning engineering, fullstack engineering, "ai engineering". primarily stuff to support ml/ai development and deployment.

some topics i have spent time on, am spending time on, or hope to spend time on include:

• data efficient finetuning, methods to produce quality synthetic and partially synthetic data, environments that simulate real world envs
• post training, sft/rl/hierarchical curriculums
• separate vs unified omni modality architectures
• audio stuff, multitask audio learning, efficient speech inference
• fast inference, io/hardware aware algos

blog

• breakdown of the grpo objective
• kv cache, post training, and hardware aware algos
• ai from first principles (talk)
• llm evalulations
• audio question answering
• engineer >> scientist
• utilizing vector embeddings
• how to contribute to open source (medium post)
• run a 46B model on your home gpu (medium post)
• train an object detection model on roboflow (medium post)

projects

• kerneltune - training models to write performant triton kernels, dataset and model published
• pip install chatan - use chatan to quickly create synthetic datasets for finetuning with llm generators and class sampling
• realtime omni - omni/multimodal inference server with qwen 2.5 omni or phi 4 multimodal. text-to-text, text-to-speech, audio-to-text, audio-text-to-text, image-text-to-text, audio-text-to-speech
• audio-llama - multimodal adapter for llama to enable paired text+audio inputs
• realtime speech-speech inference - fast stt+chat+tts
• ai notebook - cursor-esque flow in a modern notebook interface
• llm evaluator - prompt+model evals, dual chat comparison, more
• dialstructure - audio transcription, diarization, summarization, classification
• groq-nvim plugin - ai code gen in vim
• lambda labs client - manage gpu clusters with a python client