Software Engineer
ML Compilers · AI Inference and Optimization
M.S. in Computer Science, University of Southern California
AMD ROCm Certified Associate
Open Source · Compiler Engineering · AI Inference · AI Engineering · Projects · Technical Skills
| Pull request | Contribution |
|---|---|
| IREE #25089 | Encoding interfaces Generalized encoding capability checks across type propagation, HAL-to-Stream conversion, and dispatch dimension handling. |
| torch-mlir #4791 | Quantized operator lowering Lowered quantized batched matrix multiplication, handling zero points, signedness, and integer widths. |
| torch-mlir #4810 | GridSample lowering correctness Improved nearest-even rounding and tensor access safety, with numerical regression tests. |
| LLVM #221490 | MLIR control-flow correctness Fixed switch-case parsing and printing for wide and negative integers. Added range diagnostics and regression tests. |
| IREE #24953 | CPU and WebAssembly math support Added double-precision math support in IREE’s LLVM CPU backend and WebAssembly bitcode libraries. |
| IREE #24924 | AMD RDNA4 code generation Added tests for matrix instruction selection and packed tensor layouts, with CMake and Bazel integration. |
- Developed C++ compiler changes across LLVM, MLIR, torch-mlir, and IREE, covering tensor lowering to Linalg, numerical correctness, and runtime integration.
- Worked on LLVM pass plugins and runtime instrumentation, including pass registration, IR traversal, instrumentation placement, and runtime linkage.
- Validated compiler transformations and generated code with lit, FileCheck, IR inspection, numerical comparisons, and end-to-end execution.
- Built the Edge AI Inference Benchmark for vision, vision-language, audio anomaly detection, and speech recognition. Exported models to ONNX, verified inference outputs, and integrated them into ONNX Runtime.
- Prepared MMStar, ImageNet, NYUv2, ToyADMOS, and LibriSpeech for evaluation. Implemented task-specific accuracy scoring and ran inference performance benchmarks separately.
- Benchmarked vLLM serving on AMD Instinct MI300X, measuring time to first token and generation throughput.
- Fine-tuned Llama 3.1 8B Instruct on GSM8K with Unsloth, LoRA, and GRPO. Tested the saved adapter with sample inference.
- Built an AI observability assistant that connects to Prometheus, Grafana, and Datadog through MCP servers, with schema discovery, query generation, validation, and retries.
- Developed a semantic search API with FastAPI, PostgreSQL, and pgvector, using HNSW indexing, tenant isolation, and group-based access controls.
- Lightweight Multilingual Translation for Edge Devices. Distilled language-specific MarianMT teachers into a smaller multilingual student and applied post-training quantization for edge inference. Project repository
- Multimodal Stock Forecasting. Compared SARIMA, LSTM, T5, and Temporal Fusion Transformer models using historical prices, text sentiment, and market indicators. Project repository
| Area | Tools |
|---|---|
| Compiler and build tools | C++ Python LLVM MLIR Linalg torch-mlir IREE CMake Bazel lit FileCheck |
| AI inference and fine-tuning | PyTorch ONNX ONNX Runtime vLLM LoRA GRPO |
| GPU programming | HIP ROCm Triton |
| Other | FastAPI Node.js PostgreSQL pgvector Docker Kubernetes MCP |
