InSiSo Technologies Private Limited | Bengaluru

Paid internship · ₹10,000/month · Fast-track to full-time for strong performers. Must be willing to relocate to Bengaluru.

About InSiSo

InSiSo builds InSiSoNet — a proprietary, hardware-agnostic vision AI architecture that runs entirely on-device, with no cloud dependency. Our models are trained from scratch on our own architecture and deployed on constrained edge silicon, where every millisecond counts.

Bosch Rexroth ctrlX Partner · NVIDIA Inception · Intel ISV Partner · SFAL K-Tech Centre of Excellence (Govt. of Karnataka)

How this role works The internship is a working assessment, not a queue. If you're performing, we move you to full-time quickly — we're not going to make a good engineer wait out a fixed term. That's why relocation readiness matters from day one: full-time here means being in Bengaluru and on client deployments.

What you'll work on

You'll work directly on our proprietary detection models — improving accuracy, shrinking them, and making them run faster on edge hardware.

- Improve detection accuracy on industrial and logistics use cases: object detection, quality inspection, defect and anomaly detection
- Train and evaluate models on the InSiSoNet pipeline; every change measured against before/after metrics
- INT8 quantisation and quantisation-aware training — hold accuracy while cutting size and latency
- ONNX export and compilation for edge accelerators
- Analyse failure modes on real customer imagery and fix what's actually wrong, not what's easy to fix
- Contribute to our internal annotation and training tooling
- Document your work so another engineer can reproduce it without asking you

What you need

Core





- Strong grasp of object detection architectures — anchor-free heads, feature pyramids, loss design. You should be able to explain why an architecture is shaped the way it is, not just call the API.
- PyTorch at depth — custom layers, custom training loops, custom loss functions
- Model optimisation: quantisation (INT8, QAT), pruning, knowledge distillation
- ONNX export and the debugging that comes with it
- Ability to read a paper and implement it
- You can look at a PR curve or a confusion matrix and say what's actually wrong with the model

Also expected

- Python, robust. Git. Linux comfort.
- Numerical intuition — you notice when a metric is too good to be true

Good to have
- Training from scratch rather than fine-tuning pretrained weights
- Inference runtimes: TensorRT, OpenVINO, TFLite, Hailo SDK
- Efficient architecture design for constrained hardware
- C++
- Published work, Kaggle placings, or open-source contributions in CV

Who this suits Recent graduates in any computer science/AI or quantitative discipline. What matters is that you've trained a model end to end and can defend every choice in it — architecture, loss, augmentation, evaluation. Your branch matters less to us than whether you can explain why your model failed.

Working arrangement

- Starts remote or on-site depending on your situation, but you must be willing to relocate to Bengaluru . Full-time conversion means being here.
- Daily 60-minutes standup, camera on, fixed time
- Short written update at end of each day
- Travel for deployment and integration work; costs covered

Duration - 3 monthsTo apply Send your CV plus a short note covering:
1. One thing you've built. What broke the first time, and how you fixed it.
2. Which role you're applying for.

Your availability — start date, relocation readiness.

📌 Deep Learning Engineering Intern — Edge Vision (India)
🏢 InSiSo Technolgies
📍 India