Edge AI & ML Engineer. I build real-time inference systems: optimized TensorRT pipelines on NVIDIA Jetson and data-center GPUs, and LLM-powered agents that turn models into working tools.
- Hands-on with most of the NVIDIA edge lineup: Jetson Nano, the Jetson Orin family, Jetson AGX Thor and DGX Spark.
- Deploy DeepStream / GStreamer pipelines on Jetson, with one codebase for Jetson and dGPU. The platform is detected at runtime and Tegra-specific elements are added only when needed.
- Build TensorRT FP16 engines from ONNX models, directly on the target device.
- Use hardware-accelerated video through NVDEC/NVENC (
nvv4l2decoder,nvv4l2h264enc), with decoders tuned for maximum performance on integrated GPUs. - Ingest multiple RTSP cameras and run them through a single batched inference call, with live-source latency handling.
- Deploy in containers with NVIDIA's multi-arch DeepStream images.
- Built an AI job search agent that works in several steps:
- An LLM turns an uploaded CV into a structured profile and plans the searches.
- The agent queries job boards and employer ATS APIs in parallel and filters the results with rules.
- An LLM scores each posting's fit in parallel batches and writes a tailored CV and cover letter.
- Provider-agnostic LLM layer: Claude, OpenAI-compatible APIs or fully local models through Ollama. Every response is validated against a Pydantic schema.
- Serve and run LLMs and vision-language models (VLMs) with vLLM and Hugging Face. Most of this work lives in private repositories.
- Agentic coding with Claude Code in my daily development workflow.
Detection, multi-object tracking, instance segmentation and multi-person pose estimation. This includes custom post-processing written directly on raw output tensors.
| Project | What it does | Stack |
|---|---|---|
| AI Job Search Agent | Multi-step LLM agent: CV parsing, search planning, multi-source job retrieval, LLM fit scoring and tailored CV generation | Python · FastAPI · Pydantic · Next.js · Claude / OpenAI / Ollama |
| Basketball Player Tracking | Player detection and tracking with stable IDs, live per-player stats and a court minimap | DeepStream · PeopleNet · NvDCF · Docker |
| Multi-Stream Pose Estimation | Multi-person pose estimation on several RTSP/file streams, with skeletons decoded from raw tensors | DeepStream · TensorRT · GStreamer |
| YOLO Instance Segmentation | Multi-stream YOLO11 segmentation with end-to-end TensorRT engines | DeepStream · TensorRT · YOLO |
| DeepStream Pose Estimation in Python | Python port of NVIDIA's C++ pose estimation app, with PAF decoding and Hungarian matching | DeepStream · TensorRT · NumPy |
| SwiftWheels | Vehicle rental system whose business logic lives in PostgreSQL procedures and triggers | Java · PostgreSQL |
| Area | Tools |
|---|---|
| Edge AI & inference |       |
| LLMs & agentic AI |        |
| Computer vision |    |
| Languages |      |
| Backend & tools |       |