Curate the signal
Label sparse inputs, clean noisy datasets, and build repeatable data workflows when clean training sets do not exist.
Open to ML opportunities
Machine Learning Engineer · Kathmandu
Turning messy data into working intelligence.
I build and evaluate practical ML systems—from self-labeled satellite imagery to locally deployed language models—when the data is sparse, the compute is limited, and the path is not obvious.
Capabilities
From ambiguity to evidence
I work across the full ML lifecycle, with a bias toward measurable outcomes and systems that survive real constraints.
Build the model
Fine-tune LLMs, train tabular classifiers, and develop segmentation pipelines with compute constraints in mind.
Prove it works
Choose task-appropriate metrics, investigate failure modes, and deploy locally when cloud APIs are not an option.
Selected work
Built, measured, documented
01
FlyRank — Content Decline Triage
Random Forest classifier that detects observed decline and ranks editorial review opportunities across 144K content items.
0.77 ROC AUC100% P@50
See the case study
02
LLaMA-2 Fine-Tuning for Code Generation
QLoRA fine-tuning of LLaMA-2 7B Chat for practical Python generation on consumer GPUs.
+19.4% Pass@515.67 tok/s
See the case study
Let’s build something useful
Have a difficult ML problem?
I’m interested in early-stage teams where clear thinking, hands-on execution, and honest evaluation matter.