New opportunity

Computer Vision AI Intern

Zycus · Bangalore, India

About this role

Zycus’s Ataavi is seeking a Computer Vision AI Intern in Bangalore (Hybrid) to help develop bird-identification capabilities using photographs and audio recordings. The hands-on internship focuses on improving the accuracy and robustness of AI systems for a biodiversity application, with the possibility of contributing features used by birders across India. The posting lists 0–1 years of experience.

Responsibilities

Build and improve image-based bird-identification models, addressing real-world challenges such as varied angles, backgrounds, lighting and partially visible birds.
Work on bird-call and bird-song identification using audio machine learning. Experiment with spectrogram-based classification, audio embeddings, multimodal models and foundation models.
Evaluate open-source models and datasets, including approaches similar to BirdNET; design experiments, benchmark accuracy and identify performance improvements.
Potentially contribute to production features in the Ataavi birding app.

Candidate profile:

Students in any year of college may apply; demonstrated ability and curiosity matter more than year or degree. The posting ideally seeks strong interest in computer vision, machine learning or deep learning; hands-on Python experience with frameworks such as PyTorch or TensorFlow; and experience with image classification, object detection, embeddings, transformers or related techniques.
A meaningful computer vision or ML research project, paper, internship or independent project, plus the ability to read research papers and turn ideas into working experiments, is valued. Applicants who have built models, reproduced papers, worked with large datasets, entered ML challenges or pursued independent AI projects are especially suited to the role.
Exposure to audio or signal processing or audio classification is an advantage. Interest in birds, wildlife or conservation is a bonus, not a requirement. Students from leading engineering and science colleges with strong projects and research work are particularly encouraged to apply.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Computer VisionMachine LearningDeep LearningPythonPyTorchTensorFlowImage ClassificationObject DetectionEmbeddingsTransformersAudio ClassificationSignal ProcessingSpectrogram-Based ClassificationAudio EmbeddingsMultimodal ModelsFoundation ModelsModel EvaluationResearch Paper Analysis

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Improve bird identification from photographs
The role involves building image-based models that remain accurate across varied angles, backgrounds, lighting and partial visibility. Prepare to explain an image-model experiment, including how you evaluated difficult examples.
Explore bird-call and bird-song identification
Audio ML is a stated work area, with spectrogram classification, audio embeddings, multimodal models and foundation models named as approaches to explore. Prepare to explain how you would compare approaches; prior audio experience is described as an advantage, not a requirement.
Benchmark existing approaches and improve performance
The posting calls for evaluating open-source models and datasets, including approaches similar to BirdNET, then designing experiments and benchmarking accuracy. Prepare to describe a reproducible comparison and how its results would guide the next experiment.
Turn research ideas into working experiments
Zycus values meaningful computer vision or ML project work and the ability to translate papers into experiments. Prepare a concrete account of an implemented idea, the choices made and what the results showed; contribution to Ataavi app features is only a possibility.

Likely challenges to think through

Robustness beyond clean photographs
Preparation hypothesis: Improving accuracy on varied or partially visible birds may require balancing overall benchmark gains against performance on hard cases. Prepare to discuss test slices, representative failure examples and how you would check whether a change genuinely helps.
Choosing among audio and multimodal approaches
Preparation hypothesis: Spectrogram, embedding and multimodal methods may differ in accuracy, complexity and the data they need. Prepare a small comparison plan with a common evaluation set and clear criteria for choosing a next experiment.
Translating benchmarks into potential app features
Preparation hypothesis: A model that performs well in an experiment may not behave as reliably on photographs or recordings supplied by birders. Prepare to explain how you would inspect real-world errors before recommending a feature; the posting does not promise production deployment.

Practice your answers

Technical practice

How would you choose between image classification and object detection for identifying a bird in a photograph?

What to coverThe posting names both techniques and highlights partially visible birds and varied backgrounds. Preparation suggestion: Explain what each approach predicts, when locating the bird might help, and how you would compare their results on representative photographs.

How would you make an image model more robust to changing angles, lighting and backgrounds?

What to coverThese are challenges named in the posting. Preparation suggestion: Discuss relevant training data or augmentations, then explain how you would test whether improvements hold across each condition rather than only on average.

When might image embeddings, transformers or foundation models help bird identification?

What to coverThe posting names embeddings and transformers among relevant techniques and proposes experimenting with foundation models. Preparation suggestion: Compare a pretrained approach with a simpler image baseline, including the evidence you would need before choosing the more complex option.

How would you build a spectrogram-based bird-call classifier?

What to coverThe role includes bird-call and bird-song identification and spectrogram-based classification. Preparation suggestion: Walk through converting recordings into model inputs, training a classifier, and checking how background sound or variable recording length affects predictions.

How would you evaluate audio embeddings for identifying bird calls?

What to coverAudio embeddings are an approach the posting proposes exploring. Preparation suggestion: Explain how you would compare recordings using embeddings, establish an audio-classification baseline, and inspect which calls the approach confuses.

When would combining photographs and audio improve identification?

What to coverThe posting proposes multimodal models for a system that identifies birds from photographs and recordings. Preparation suggestion: Describe how you would compare image-only, audio-only and combined predictions, including what to do when one input is absent or unreliable.

How would you assess an open-source biodiversity model or dataset before using it?

What to coverThe role calls for evaluating open-source models and datasets, including approaches similar to BirdNET. Preparation suggestion: Examine task fit, species coverage, data quality and evaluation setup; distinguish published results from performance measured on the intended data.

How would you benchmark identification accuracy and diagnose weak results?

What to coverCreating experiments, benchmarking accuracy and finding improvements are stated duties. Preparation suggestion: Describe a held-out evaluation, appropriate class-level measures and error analysis by species and recording or photograph conditions, while watching for data leakage.

How would you keep Python model experiments comparable?

What to coverThe posting seeks hands-on Python experience with frameworks such as PyTorch or TensorFlow. Preparation suggestion: Explain how you would hold data splits and evaluation criteria consistent while recording configurations, model versions and results across experiments.

What would you check before proposing a model improvement for the Ataavi birding app?

What to coverThe posting says an intern might contribute to production features, without specifying deployment requirements. Preparation suggestion: Explain how you would verify an accuracy gain on realistic inputs and, if considering an app feature, investigate practical constraints rather than assume them.

Behavioural practice

Tell me about a computer vision or ML project you pursued beyond standard coursework.

What to coverThe posting values meaningful projects and independent experimentation. Preparation suggestion: Use a real situation, the specific work you chose to do, and the measured result or lesson; be clear about your own contribution.

Describe a time you turned a research-paper idea into an experiment.

What to coverReading papers and translating ideas into working experiments is explicitly valued. Preparation suggestion: Give the situation, how you selected and implemented a testable idea, and what the result showed, including any gap between the paper and your reproduction.

Tell me about an experiment that did not improve a model as expected.

What to coverThe role focuses on experiments and improving accuracy and robustness. Preparation suggestion: Explain the initial objective, how you diagnosed the result and what you changed or concluded; report the outcome without claiming a gain that did not occur.

Describe a time you had to learn an unfamiliar ML technique to address a problem.

What to coverThe internship spans computer vision and audio ML, with audio exposure described as an advantage rather than a requirement. Preparation suggestion: Choose a real example, explain how you learned and tested the technique, and state the result or limitation.

Sharpen your knowledge

Topics to prepare

Image classification, detection and embeddings
The posting seeks experience with these techniques for bird identification from photographs. Revision exercise: explain when each approach could help identify a bird that occupies only part of an image, and what each approach might miss.
Visual robustness
The role explicitly addresses varied angles, backgrounds, lighting and partially visible birds. Revision exercise: sketch a test set divided by those conditions and predict where an image model might fail.
Bird-audio representations
The work includes bird-call and bird-song identification, with spectrogram classification and audio embeddings named as approaches. Revision exercise: compare how each represents a recording and how background sound could affect it.
Multimodal and foundation-model approaches
The posting names these as approaches to experiment with. Revision exercise: outline a possible comparison between image-only, audio-only and combined predictions; treat the design as preparation, not a specified implementation.
Experimental evaluation
The role calls for evaluating open-source models and datasets, benchmarking accuracy and finding improvements. Revision exercise: design a comparison that uses the same held-out examples for two models and reports both overall accuracy and errors by challenging condition.

Skills to revise

Build an image baseline in Python
Hands-on Python and a framework such as PyTorch or TensorFlow are listed as ideal experience. Practice implementing a small image-classification baseline, recording its data split, predictions and accuracy for a bird-identification example.
Inspect difficult image predictions
The posting emphasizes real-world photographs. Practice sorting model mistakes by angle, background, lighting and partial visibility, then test one change and report whether it improves the relevant group.
Prototype an audio classifier
Audio classification exposure is an advantage, and spectrogram-based classification is named work. As preparation, turn labeled bird-call clips into spectrogram inputs, run a simple classification experiment and inspect incorrect predictions.
Reproduce and benchmark a published approach
Reading papers, turning ideas into experiments and evaluating open-source approaches similar to BirdNET are explicitly relevant. Practice extracting one testable idea from a paper, implementing or adapting it, and comparing results against a baseline on the same dataset.

Your preparation checklist

Prepare a project walkthrough
Choose a computer vision or ML project, if you have one, and explain your contribution, Python and framework choices, experiments, results and limitations. The posting values meaningful projects and hands-on work with PyTorch or TensorFlow.
Sketch an image-identification experiment
As interview practice, outline a bird-image baseline and tests for varied angles, backgrounds, lighting and partial visibility. Explain how you would inspect errors and decide what to improve.
Outline an audio approach
Prepare to explain how you would test bird-call or bird-song identification, for example with spectrogram classification or audio embeddings. Treat this as a proposed approach, not a claim of audio experience.
Draft a model comparison plan
Create a small comparison template for open-source models and datasets, including BirdNET-like approaches: data splits, accuracy measures, failure cases and how you would make results comparable.
Show how you turn research into experiments
Select a relevant paper you have read and summarize its idea, a feasible reproduction test and what result would challenge your hypothesis. The posting explicitly values turning papers into working experiments.
Prepare questions about the work
Ask which image and audio failure cases matter most, how model improvements are evaluated, and what would make an experiment useful for a potential Ataavi app feature.

Availability last checked: 4 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in India. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…