Subtain Malik

AI/ML Solutions Architect · Netherlands
Production AI, from the GPU to the edge.

My models work at

Fig. 01 · Procedural silicon

Sigrow · the platform

One platform.
Greenhouse to GPU and back.

Consulting · Rapidev Group and OwlSense · 2023–2025

One serving layer.
Less cloud, faster models.

DARVIS · OmniRoom · 2021–2024

Every angle of the room.
And the models keep learning.

Pakistan Air Force · AIRAF · 2020–2021

The picture is restricted.
The pipeline isn't.

Less memory. Less waiting.
More from every GPU.

88%

Less GPU memory

Peak VRAM on the same workload. A recurring crash, gone.

21.8×

Faster p95 latency

In the time one request took, 21 now finish.

23×

Faster LLM inference

Fine-tuned and quantized: 23 answers in the time of one.

$50K

Saved a year in cloud

One shared serving layer for 100K+ requests a day.

Systems I designed.
Greenhouses, feeds, cameras, satellites.

01
Inference optimizationML PlatformHortiEngine20 models on shared GPUs: 8 million inferences a day, p95 cut from 27.2 s to 1.25 s.
02
MultimodalLLM agentsOwlSenseSocial media analysis. Multimodal extraction and a multi-agent harness, all on one message queue.
03
Video analyticsML PlatformOmniRoom100+ cameras. Parallel models into one decision layer, auto-annotation into training.
04
Satellite imageryMission-criticalAIRAF450-megapixel satellite images, detected in patches; each crop matched against a vector store.