Fawad S.

Edge AI and computer vision deployment on NVIDIA Jetson, RK3588, and embedded hardware. Your model works in the notebook. On the device, it misses its deadline. That gap is where I work. I'm a Top Rated engineer with an MS in Artificial Intelligence Engineering and active research in efficient vision-model deployment on embedded hardware. I get vision models to meet real latency, power, and thermal budgets on constrained devices, and I prove it with measurements taken on your target hardware, not estimated from spec sheets. Most performance problems I'm brought in for aren't model problems. They're measurement problems: teams size hardware from FLOPs or spec-sheet TOPS, and real silicon disagrees, sometimes by more than a factor of two on the configurations I've measured. I start by measuring what your system actually does, then fix what the numbers point at. ▪️ EDGE AI & COMPUTER VISION 🎯 Hardware selection, vendor-neutral. Which board, which accelerator, which precision, for this workload and this budget? I take no margin on hardware, which is the only reason the recommendation is worth anything. If the cheaper option wins, that's what the report says. 🔍 Feasibility & technical assessment. Can this model run on this hardware, at this accuracy, inside this budget? I answer with measured evidence before you commit to a board, a sensor, or an architecture. It's the cheapest engagement you'll run, and it prevents the most expensive mistakes. 📊 Measurement & benchmarking. Worst-case latency (p99, not averages), power draw, thermal derating under sustained load, and where your pipeline actually spends its time. Locked clocks, warm-up discarded, medians over hundreds of runs. This is the protocol real deployments require. ⚙️ Deployment engineering. ONNX export with numerical-parity verification, engine building and optimization, pipeline restructuring, quantization and precision selection, memory and bandwidth tuning. From notebook to production on the hardware you're actually shipping. 🧠 Model development under constraint. Building or adapting detection, classification, and tracking models to fit a budget measured first, rather than training something large and hoping it fits later. ▪️ PLATFORMS NVIDIA Jetson (Nano, Orin Nano, Orin NX) with TensorRT and CUDA. This is where my research sits and where most of my measured deployment work has been done. Rockchip RK3588 and RK3588S boards (Forlinx, Radxa, Orange Pi) with the RKNN toolchain. I've written the bring-up and verification methodology for this family, including the silicon-variant and toolchain-host constraints that catch teams out. Raspberry Pi 4, 5 and CM4/CM5 modules, including Pi paired with an NPU accelerator where the SoC alone isn't enough. Also: Hailo, Coral Edge TPU, OpenVINO targets, other ARM SBCs, and small x86 plus discrete GPU edge boxes. On a platform I haven't measured yet, the approach is identical. The methodology is platform-neutral; the numbers never are. Experience tells me where to look, your hardware tells me what's true. Real-time detection, tracking and perception on optical and thermal (EO/IR) imagery. Offline and connectivity-denied deployments, SWaP-constrained systems, air-gapped and regulated environments. I test on real edge hardware, so I catch what cloud-only engineers miss: thermal throttling after twenty minutes in an enclosure, brownouts under inference load, preprocessing mismatches that silently cost accuracy, engines that don't transfer between machines. ▪️ RECENT CLIENT WORK ✅ Assessed an edge vision system before a 10-to-100+ unit rollout, identifying production risks that would have surfaced as field failures across the fleet ✅ Audited a PyTorch/C++ inference pipeline on embedded hardware and isolated the real bottleneck, which was not where the team had been optimising ✅ Defined the production architecture and MLOps roadmap for a new AI product, including the hardware decision ▪️ HOW I WORK I treat your project as an engineering problem: clear about what's achievable, honest about trade-offs, and specific about what I measured versus what I estimated. If something won't work on your hardware, you'll hear it early, and that is usually worth more than the work itself. Every engagement ships with its evidence: the raw measurements, the conditions they were taken under, and what I could not measure. You should be able to check my numbers, not just trust them. For larger engagements I scale delivery with a small specialist team, and can contract through our UK affiliate where a UK entity is required. 🛠️ Tools: PyTorch, TensorRT, ONNX, RKNN, CUDA, C++, OpenCV, NVIDIA Jetson, Rockchip RK3588, Raspberry Pi, Embedded Linux, Python, Docker If you're scaling a vision system, choosing hardware for a deployment, or holding a benchmark table that doesn't look right, tell me the workload and the target hardware and I'll tell you what I'd measure first.