MLOps Lead — Edge AI

The mission

A model that is accurate in a notebook and too slow, too hot or too expensive on the device has not shipped. You own the path from a trained model to a fleet running it in the field: optimisation onto edge accelerators, the compute you avoid spending by using motion instead of fighting it, the stream budget each device can carry, and remote updates that never brick a unit somebody is depending on. Cost per stream is a number you will care about personally.

The stack you'll live in

What you'll own

• Optimise and deploy computer-vision models onto edge accelerators — TensorRT, quantisation, pruning, and the NPU / ISP / VPU blocks the SoC actually gives you.
• Cut the compute before the model ever sees it: motion-optimized embeddings and algorithms, compressed-domain motion vectors, ROI and event-triggered inference — so we are not paying to re-infer a scene that has not changed.
• Own cost per stream: how many channels one device carries, at what latency, what power draw, and what rupee figure.
• Own encoding and transport end to end — H.264 and H.265 tuning, and the RTSP, RTMP, SRT, SFTP and SMTP paths — plus the bitrate and channel arithmetic behind every deployment plan.
• Build the remote update path — signed OTA, A/B partitions, staged rollout and a rollback that cannot brick a deployed unit.
• Keep it conformant — DPDP obligations, retention, and what is permitted to leave the device at all.
• Choose the silicon with the Hardware Lead, on evidence: benchmarks, thermals, and total cost — not on a datasheet claim.

What we're looking for

• 6+ years shipping ML on edge or embedded hardware in production, with real fleet-scale exposure.
• Deep with TensorRT / ONNX / CUDA and at least one accelerator family — Jetson, Hailo, Ambarella, Rockchip or similar.
• Real command of video codecs — H.264 and H.265 — and of using motion rather than fighting it: motion vectors, temporal redundancy, motion-optimized embeddings and algorithms.
• You understand on-chip peripherals — ISP, NPU, VPU, MIPI-CSI, DMA — and what each one costs you in practice.
• Fluent in the transport layer: RTSP, RTMP, SRT, SFTP and SMTP, and clear on when each is the right answer.
• Comfortable across CPU architectures (ARM, x86) and clear on what each implies for the build and the bill.
• You have built or run a real OTA / firmware-update system — including the rollback, which is the part that matters.

Signals that set you apart

• Cost-per-stream or power-per-stream optimisation you can put an actual number on.
• Compressed-domain or motion-vector work, adaptive sampling, or anything that made inference cheaper without making it worse.
• DPDP / GDPR-style compliance inside a device product, or government and public-safety deployments.

Who you are — a Pinacan

How we hire for this role

These are the actual rounds — no surprise stages. We reply to everyone.

Apply