Working on ML infrastructure and developer tooling

Extending the LTX text-to-video pipeline with image-to-video generation, enabling video synthesis conditioned on an input image and text prompt. Implementing full C++ and Python API support, throughput/latency benchmarking, and a test suite ensuring consistency across both implementations.

Working on bringing the OpenVINO backend for Keras 3 to functional parity with the other backends — implementing operators that were missing for OpenVINO, cutting down excluded tests along the way, and now extending the same work to keras-hub. Also reported a handful of upstream bugs in OpenVINO found while doing this.

Added LoRA adapter support for GGUF-format models, extending what was previously safetensors-only, and brought LoRA support to the Text2Video (LTX-Video) pipeline, including fixes for adapter persistence across model reshapes and a Python API for setting adapter. Also added per-stage latency metrics for the LLM and Whisper pipelines, and fixed a missing opset re-export bug.

Added environment variable support to git-clang-format for incremental formatting in CI, so large codebases can adopt formatting without noisy full-file diffs. Started as a small standalone tool, later contributed upstream into LLVM.

A web tool that checks whether a given HuggingFace model will fit on your hardware — estimating storage and VRAM across quantization formats (FP32, FP16, INT4, GGUF, and others) and checking compatibility against consumer GPUs, cloud instances, and Apple Silicon before you download anything.

FastAPI · PyTorch · FAISS · LangChain A retrieval-augmented generation system for querying documents — semantic search via FAISS, context-aware answers via LLMs, and some work on making it fast (async + batching) and reliable under load.