Subtain Malik
Senior AI/ML Engineer · Netherlands
Journey
Hover a point to see the full storyTap a point to see the full story (Work experience Other achievements Current role)
-
2018
Gold MedalMS, NUST · 2018 – 2021
MS Computational Science & EngineeringNUST · 2018 – 2021
- Graduated top of the programme with the Gold Medal
- BS Computer Science, University of Gujrat · Top Project award
-
2020
Pakistan Air Force2020 – 2021
Machine Learning EngineerIslamabad
- AI detection on 450-megapixel satellite images
- Mission-critical AI analytics infrastructure
-
2021
DARVIS2021 – 2024
Machine Learning EngineerUnited States, remote
- Real-time video analytics, +50% FPS with custom DeepStream plugins
- One NVIDIA Triton model store for PyTorch, TensorFlow and TensorRT
- Automated path from training to TensorRT to API
-
2022
Code in open sourceTensorRTX · PytorchX · 2022
Merged upstreamTensorRTX · PytorchX · 2022
- ArcFace face recognition on TensorRT 8, in TensorRTX
- MLP example networks in TensorRTX and PytorchX
- PyCon speaker
-
2023
Published researcharXiv · 2023
Reimagining UI Design using Deep LearningarXiv · 2023
- Challenges and opportunities of deep learning for app interface design
- ICT Endowment Research & Development Funds award
-
2023
AI/ML Consultancy2023 – 2025
AI/ML ConsultantSelf-employed
- LangGraph multi-agent systems on AWS Bedrock
- LoRA/QLoRA fine-tuning, 23× faster LLM inference
- One serving layer for over 3 million requests a month, $50K a year saved
-
2024
Sigrow2024 – now
Senior Software Engineer, AI/MLNetherlands
- Owns the ML serving and data platform
- 20 vision models, 8 million predictions a day, p95 27.2s → 1.25s
- Real-time data from 5K+ IoT devices
Career highlights
-
01
21.8×
Inference optimization
Cut p95 latency in production from 27.2 seconds to 1.25.
-
02
8Million
Predictions a day
20 vision models on shared GPUs, serving 500+ devices in the field.
-
03
88%
Less GPU memory
Cut peak VRAM on the same workload and ended a recurring crash.
-
04
$50K
Saved every year
One shared serving layer for over 3 million requests a month, on far less cloud compute.
-
05
+50%
More video per GPU
Custom DeepStream plugins raised real-time video analytics throughput.
-
06
450MP
Seen from orbit
AI detection on huge satellite images, 15,000 × 30,000 pixels each.