AI/ML Engineer Intern
Work on the NLP system behind Nepal’s emerging review-trust infrastructure by developing NLP systems that detect fake, manipulated, and coordinated reviews in English-language data.
- Department
- Technology
- Location
- Kathmandu, Nepal
- Workplace
- Hybrid
- Employment type
- Internship
We are building something new for Nepal: a review-trust system designed around the realities of Nepalese businesses and online users. Our core challenge is identifying whether a review reflects a genuine customer experience or an attempt to manipulate a business's reputation.
As an AI/ML Engineer Intern, you will work primarily on the fake-review detection engine, with a strong focus on NLP, text classification, representation learning, anomaly detection, and model evaluation.
The data will not behave like a clean Kaggle dataset. You will work with inconsistent writing styles, spelling variations, emojis, copied or templated phrases, unusually repetitive reviews, suspicious activity patterns, and increasingly sophisticated attempts to make fake reviews look human. The challenge is not simply achieving a high score in a notebook. It is understanding what trustworthy review behavior looks like in the Nepalese market and turning those signals into a system that can be measured, improved, and eventually deployed.
You will have meaningful ownership from the early stages of the product. Depending on your strengths, you may help define labeling strategies, establish the first model baselines, experiment with transformer-based approaches, investigate difficult false positives, design evaluation datasets, and work with the engineering team to bring successful models closer to production.
This role is for someone who enjoys asking “Why is the model getting this wrong?” rather than only asking “How do I increase the accuracy?”
What you’ll do
- Build and iterate on NLP models for fake, suspicious, and low-authenticity review detection.
- Work primarily with English-language review data from real-world business and customer interactions.
- Build strong baseline models and experiment with transformers, embeddings, classical machine learning, and deep learning approaches.
- Investigate linguistic patterns associated with fake, copied, templated, incentivized, or coordinated reviews.
- Explore signals such as similarity, repetition, semantic inconsistency, unusual phrasing, review structure, and behavioral patterns.
- Design experiments that account for class imbalance, noisy labels, data leakage, distribution shifts, and adversarial behavior.
- Build reliable data-processing and feature pipelines for training and evaluation. Perform detailed error analysis to understand why models generate false positives and false negatives.
- Perform detailed error analysis to understand why models generate false positives and false negatives.
- Evaluate models using metrics such as precision, recall, F1, ROC-AUC, PR-AUC, calibration, and threshold-based performance, rather than relying on accuracy alone.
- Help create and continuously improve a high-quality labeled dataset for review authenticity.
- Research relevant work in NLP, spam detection, anomaly detection, misinformation, trust & safety, and content authenticity.
- Turn experiments into reproducible training and evaluation workflows instead of isolated notebooks.
- Work with software engineers to move promising models from research into production.
- Document experiments, assumptions, failures, and findings clearly so the team can build on them.
What we’re looking for
- Strong fundamentals in Python and practical experience with tools such as PyTorch, TensorFlow, scikit-learn, pandas, or Hugging Face.
- Genuine interest in machine learning and NLP, demonstrated through projects, coursework, research, or serious independent work.
- Good understanding of core ML concepts including classification, embeddings, train/validation/test splits, overfitting, regularization, and model evaluation.
- Comfortable working with messy real-world text and investigating the data instead of expecting perfect inputs.
- Able to read technical papers and turn useful ideas into practical experiments.
- Strong analytical thinking and curiosity about why models fail, not just whether they work.
- Familiarity with transformers, LLMs, text embeddings, semantic similarity, or modern NLP techniques is a strong advantage.
- Experience with text classification, spam/fraud detection, anomaly detection, recommendation systems, or trust and safety problems is a plus.
- Familiarity with Git, Linux, APIs, data pipelines, or basic model deployment is helpful.
- Strong communication and documentation habits.
- Willingness to learn quickly, experiment independently, and work through problems that do not already have obvious answers.