Machine Learning Engineer — Computer Vision

Lear Labs; Bengaluru; On-site; INR 3,500,000 - 7,000,000; Full time

Benefits offered by this job

Real ownership of production systems

Direct founder access

Strong hardware resources

Job summary

Lear Labs in Bengaluru, India, is building a production computer-vision platform handling large volumes of imagery and video from moving devices. You will own model architecture, data pipelines, labeling standards, evaluation harness, and GPU infrastructure, with growth into a ML lead role over time.

You will work on detection, tracking, data management, and inference at scale, ensuring reliability, cost efficiency, and performance for non-technical decision-makers through dashboards and reports.

Qualifications

• Hands-on computer vision work with real users
• Fine-tuning object detection on custom data (YOLO or equivalent)
• Solid PyTorch and CNN fundamentals
• Judgement about data; fix data issues affecting models
• Comfortable on Linux with Docker and cloud GPUs
• Excellent written/spoken English and basic Hindi
• Strong track record from collegiate CV projects or real deployments

Responsibilities

• Develop and own detection models and thresholds
• Implement tracking, de-duplication, and counting across frames
• Design and manage training data pipelines and labeling standards
• Handle video/image ingestion, frame sampling, and quality controls
• Evaluate models with per-class metrics and regression checks
• Scale data/infrastructure across GPUs and multi-node setups
• Maintain and optimize the training environment and GPU stack

Skills

Hands-on CV

Fine-tuning detection

PyTorch fundamentals

Data judgement

Linux & Docker

English + Hindi communication

Collegiate CV projects

Tools

Docker

Linux

Cloud GPUs

PyTorch

Job description

About Lear Labs

Lear Labs is an enterprise AI lab dedicated to solving the hardest AI and machine learning problems for leading organizations in India and beyond. We move past theoretical proofs-of-concept to build robust, production-ready systems. Our partners include top-tier enterprises, sprawling conglomerates, and government bodies who rely on us to deliver models that work reliably in the real world.

About the Role

We are building a production computer-vision platform that ingests large volumes of continuous imagery and video captured from moving devices operating at variable speeds. It turns this data into structured, geo-referenced findings, and serves those findings to non-technical decision-makers through dashboards and reports.

This is not a research seat and it is not a proof-of-concept. The models you build will run continuously against real-world footage — shot at speed, in bad light, on cheap lenses, across seasons — and the output will be used by people who make budget and operational decisions on it. Accuracy, reliability, and cost per processed hour of video all matter.

You will be the technical owner of the vision and data side of the platform: model architecture, the training data pipeline, the labelling standard, the evaluation harness, and the GPU infrastructure it all runs on. As the platform grows, this role grows into an ML lead position with a team under it.

What the Work Involves

• Detection models. Fine-tuning object detection models on custom data — including classes that are small, low-contrast, or partially occluded. Choosing thresholds against a real objective rather than a leaderboard metric.
• Tracking and counting. An object appearing across a hundred frames is one object, not a hundred. Tracking across occlusions and gaps, de-duplication, and counts that hold up when someone checks them by hand.
• Training data. This is where most vision projects actually fail. Defining what gets labelled and how, running annotation quality checks, handling class imbalance and long-tail cases, deciding what to label next, and keeping datasets versioned so any model can be traced back to what trained it.
• Video and imaging. Frame sampling from continuous video, and understanding how frame rate, resolution, exposure, motion blur, lens distortion, and compression affect what a model can and cannot see. Being able to say what a good capture looks like instead of accepting whatever arrives.
• Evaluation. A single mAP number is not an evaluation. Per-class precision and recall at chosen thresholds, performance sliced across the conditions that vary in data, frozen test sets, and regression checks before anything ships.
• Data and infrastructure at scale. Terabytes of images and video with metadata attached to every frame. Storage layout, indexing, and pipelines that keep GPUs busy rather than waiting on I/O. Multi-GPU training in the cloud, reproducible pipelines with experiment tracking, and inference that scales predictably.
• Your own environment. Linux, Docker, the CUDA and driver stack, and secure SSH access. When the driver stack breaks, it is yours to fix.

What We Are Looking For

Must have — the fundamentals

• Hands-on computer vision work, including at least one system that reached real users and that you kept working on afterwards. Years matter less to us than what you have actually shipped.
• Fine-tuning object detection models on custom datasets — YOLO family or equivalent.
• Solid PyTorch and CNN fundamentals. Comfortable reading and modifying a training loop, not only calling a wrapper library.
• Judgement about data. You can look at a training set and say what is wrong with it, and you have fixed a data problem that was pretending to be a model problem.
• Comfortable on Linux with Docker and cloud GPUs. You can set up your own training environment and keep it running.
• Clear communication in English and Basic spoken in Hindi. You will write things down — guidelines, evaluation notes, what you found — and explain limitations honestly to people without an ML background. You raise problems early rather than late.
• Projects. If you have participated in collegiate competitions—such as SAE (BAJA, Formula Student / FSAE), robotics challenges, or high-stakes computer vision hackathons—or built hands-on systems from scratch, make sure to highlight this upfront. We care deeply about real-world engineering grit over theoretical research.

Strong plus — but learnable here if the fundamentals are there

• Multi-object tracking and re-identification.
• Vision Transformers, and a view on when they are worth the compute and when they are not.
• Working with large datasets end to end: ingestion, cleaning, de-duplication, sharding, versioning.
• Segmentation, or estimating size and extent rather than just drawing a box.
• Inference optimisation — ONNX, TensorRT, quantisation, deployment to edge or mobile hardware.
• Experiment tracking and MLOps tooling (Weights & Biases, MLflow, DVC, or equivalent).
• Networking and access setup: bastion hosts, VPNs, firewall rules, key management.
• Model monitoring and drift detection in production.
• Self-supervised or semi-supervised pre-training on unlabelled data.
• Prior work in domains where imagery is captured in uncontrolled outdoor conditions — infrastructure inspection, agritech, autonomous driving, industrial QC, drone or aerial imagery.

You are not expected to have all of the second list. Two or three, plus the fundamentals, is a strong application.

What you'll Get

• Real ownership. Features, not tasks. Production systems, real users, actual stakes.
• Direct founder access. Work closely with the CEO and CTO on architecture, product, and technical strategy.
• Accelerated growth. Hard problems and a tight feedback loop — you will grow more in six months here than in two years most places.
• Serious hardware. Access to powerful compute for machine learning work.
• Depth ahead of you. The vision and ML work here is not a single build with an end date.
• Compensation. Competitive and negotiated case-by-case. We evaluate compensation individually based on your depth of experience, technical capability, and what you bring to the table. We discuss expectations openly during the initial conversation.

Location

We are based in Bangalore. You will also travel to our sites in Bhopal for active fieldwork and working at our operations hub. We treat data collection as a first-class engineering problem. You will travel to the field to see exactly how your footage is captured on the ground, because understanding the physical reality of the data pipeline is the only way to build models that actually survive in production.