Data Science Intern — Computer Vision & Agentic AI

About CerebralZip:

We exist to accelerate India's presence in cybersecrutiy innovation by building Gen AI based products in All Source Intelligence (ASI). Founded in January 2024, we have successfully built and deployed one of product pertaining to digital forensics in a big enterprise organisation.

Being first of its kind in India, we are looking to hire and groom the most energetic folks to help us take our innovative solutions to the world. Come build and scale with us.

About the Role

We are looking for a Data Science Intern who is equally comfortable training a vision model and wiring it into an agentic system. You will work on real production problems in computer vision — face recognition, ANPR, activity and crowd analytics — and on the agent harnesses that turn those models into autonomous, tool-using workflows.

This is a hands-on role. You will be writing training code, debugging data pipelines, and shipping models into containers — not just running notebooks.

What You'll Do

• Train, fine-tune, and evaluate vision models for tasks such as face recognition, activity detection, ANPR, heatmap-based tracking, and crowd density estimation.
• Build and maintain image and video processing pipelines using OpenCV — preprocessing, augmentation, frame sampling, tracking, and post-processing.
• Design and iterate on agentic AI workflows: tool calling, planning loops, memory, and multi-step task execution.
• Work with agentic harnesses to orchestrate models and tools, and to evaluate agent behaviour systematically.
• Run experiments end to end — dataset curation, annotation review, training runs, ablations, and error analysis.
• Optimize models for inference (quantization, ONNX/TensorRT export, batching) and help deploy them behind APIs.
• Containerize services with Docker and expose model inference through FastAPI endpoints.
• Document experiments, metrics, and findings clearly so results are reproducible.

Must-Have Skills

Computer Vision

• Strong hands-on experience with OpenCV for image and video processing.
• Demonstrated experience training and fine-tuning vision models in at least one of: face recognition, activity/action detection, ANPR, heatmap-based tracking, or crowd density detection.
• Solid understanding of CNNs, object detection architectures (YOLO / Faster R-CNN / DETR family), and evaluation metrics (mAP, IoU, precision/recall, ROC).

Deep Learning Frameworks

• Working proficiency in PyTorch and TensorFlow — custom datasets and dataloaders, training loops, loss functions, transfer learning, checkpointing.

Agentic AI

• Understanding of agentic AI concepts: tool use, function calling, planning and reflection loops, memory, and multi-agent coordination.
• Exposure to at least one agentic harness / framework (e.g. LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, or a custom in-house harness).
• Ability to write and iterate on prompts, and to evaluate agent outputs beyond eyeballing them.

Engineering Basics

• Comfortable with Python (clean, modular code — not just scripts).
• Familiarity with Docker (building images, running containers, basic compose).
• Familiarity with FastAPI or a similar framework for serving models.
• Git-based version control and collaborative workflows.

Good to Have

• Experience with edge deployment (Jetson, Raspberry Pi) or real-time video stream processing (RTSP, GStreamer).
• Exposure to model optimization: ONNX, TensorRT, OpenVINO, pruning, quantization.
• Familiarity with vector databases and RAG pipelines.
• Experiment tracking tools (Weights & Biases, MLflow, TensorBoard).
• Annotation tooling (CVAT, Label Studio) and synthetic data generation.
• Cloud exposure (AWS / GCP / Azure) and basic CI/CD.
• Public GitHub repos, Kaggle work, or published papers/blogs in CV or agentic AI.

Who You Are

• Pursuing or recently completed a degree in Computer Science, Data Science, Electrical Engineering, or a related field.
• You debug rather than guess — you read stack traces, inspect intermediate tensors, and visualize what the model is actually seeing.
• You can take an ambiguous problem statement and turn it into an experiment plan.
• You communicate results honestly, including the ones that didn't work.

What You'll Get

• Ownership of real features that ship, not a side project shelved after the internship.
• Direct mentorship from senior ML and engineering leads.
• Access to GPU compute and production-scale datasets.
• Certificate, letter of recommendation, and a pre-placement offer for strong performers.

In your application, briefly describe one vision model you trained or fine-tuned yourself — what the dataset was, what broke, and how you fixed it.