Software engineering graduate (B.E., Nepal College of Information Technology, 2026) based in Kathmandu, Nepal. I build machine learning systems end to end, from data pipeline to deployed API, and test them on data they have never seen: a later month, a new camera, a different site. When a model misses its target, the README says so.

Open to Machine Learning Engineer and Applied ML Engineer roles, remote or on-site.

^{NLP  ·  information extraction  ·  Live demo}

Turns free-text lab protocols into an editable preparation sheet, with every extracted value linked back to its source text. A SciBERT span-and-relation model reaches 0.766 entity F1 and 0.580 end-to-end relation F1 on WNUT-2020, ahead of a CRF baseline at 0.504.

_{PyTorch Transformers ONNX Runtime FastAPI Docker}

^{Tabular ML  ·  calibration  ·  Live dashboard}

Scores card-not-present transactions with a calibrated fraud probability and picks approve, review or block within an analyst team's daily review budget. It reaches 0.907 ROC-AUC on a held-out later month, at 43% lower cost than approving everything.

_{XGBoost scikit-learn SHAP FastAPI MLflow}

^{Computer vision  ·  image classification  ·  human-in-the-loop}

A review inbox for trail-camera photos that groups shots into capture events and suggests a species for each. Grouping alone cuts the review queue by 61.4%, and a fine-tuned EfficientNet-B0 scores 0.447 macro-F1 on nine unseen cameras, against 0.285 for the frozen baseline.

_{PyTorch FastAPI Streamlit PostgreSQL Redis AWS}

^{Computer vision  ·  detection and tracking  ·  sonar video}

Counts fish moving upstream and downstream in sonar video and keeps the evidence so a technician can check every count. It reaches 0.213 count error (nMAE) across four unseen locations, against 0.388 for a tuned classical baseline.

_{PyTorch YOLOX OpenCV ONNX Runtime FastAPI}

^{Tabular ML  ·  self-collected data  ·  Live app}

Predicts which job postings will be taken down within seven days, trained on a panel I scraped from seven job boards. On held-out days, 75% of its daily shortlist were real removals, 6.8 times the hit rate of reading the board unaided.

_{scikit-learn XGBoost FastAPI Streamlit MLflow}

^{Computer vision  ·  semantic segmentation  ·  in development}

Segments cardboard, plastic and metal contamination on paper-recycling conveyor lines and queues uncertain images for review. SegFormer-B0 reaches 0.336 foreground macro IoU, against 0.152 for a tuned random forest.

_{PyTorch SegFormer DVC MLflow FastAPI}