Production AI, shipped — not slides.
From notebook to deployed device. Computer vision, edge AI, ML pipelines, and the messy hardware ↔ model ↔ deployment work that gets a real product into customers' hands.
What I do
Edge AI Audit
One-week deep dive. Latency, memory, and quantization profile of your model on your target hardware. Three prioritized speedups with expected impact.
Notebook → Device
Take a working model and ship it to your hardware. Quantization, ONNX/TFLite/Coral export, inference pipeline, benchmark suite, handover docs.
Vision Pipeline Build
End-to-end custom CV pipeline from data spec to deployed device. Data labeling strategy, model training, ablations, edge deployment.
Fractional CV / ML Lead
Two days a week embedded with one team. For seed-stage hardware startups not yet ready for a full-time hire. Retainer.
Workshops for engineering teams
1–2 day on-site or virtual workshop. Production AI, model compression, edge deployment, real failure modes — tailored to your stack.
Working sessions
Single 90-minute paid sessions for founders or CTOs stuck on a specific model, deployment, or roadmap call. Cheapest sanity check available.
Who this is for
-
Hardware / IoT companies running CV at the edge
Smart cameras, retail analytics, agritech, industrial inspection, manufacturing QC. You have hardware engineers — not a CV+edge specialist on staff.
-
Sports-tech and analytics startups
Ball tracking, player analytics, broadcast graphics, on-court automation. (Disclosure: I'm a cofounder at a sports-tech company, so direct competitors in that specific niche are off-limits by policy. We'll establish fit on the discovery call.)
-
Robotics and embedded AI teams
Pick-and-place, AMRs, drones. Vision pipelines that have to run on Jetson Orin Nano, Coral, or RPi CM5 with real latency budgets.
-
Companies with stuck POCs
"It works on a server, but we can't ship it to the device." This is exactly where I live.
Areas I work in
The production capabilities this practice draws on:
-
Real-time object & trajectory tracking on edge devices
Single-shot detectors, per-frame heatmap models, temporal smoothing. Tuned for ARM-class hardware where every millisecond costs power and every megabyte costs throughput.
-
Action and event segmentation in video
Temporal convolutional and transformer architectures over per-frame features. Fine-tuned for in-domain footage where generic open-source models break.
-
Multi-camera calibration & 3D reconstruction
Intrinsics, extrinsics, homography, physics-aware trajectory fitting. The math that turns 2D pixel coordinates into real-world measurements you can act on.
-
Model compression & on-device deployment
Knowledge distillation, quantization (FP16 / INT8), pruning. TFLite, ONNX Runtime, Coral USB TPU, ARM NEON. Benchmarked on the actual hardware, not on a workstation.
-
Data pipelines for in-domain CV
Active learning to prioritize labels where the model is most uncertain. Semi-supervised pseudo-labeling to multiply effective dataset size. Inter-annotator quality tracking.
How an engagement works
-
01
30-minute discovery call
Free. Tell me what you're building, what's stuck, and what hardware you're targeting. I'll tell you honestly whether I'm the right person for it.
-
02
Written scope & quote
Within 3 working days. Fixed scope, fixed price, milestones, deliverables. No hourly billing on productized packages.
-
03
Delivery in milestones
Slack/WhatsApp updates, weekly demo, transparent progress against the milestone plan. You can stop at any milestone.
-
04
Handover
Code, models, benchmark suite, documentation, and a written handover doc. Optional 2-week post-handover support included.