Machine learning · Computational neuroscience
Satish Bajagain
I’m a full-stack and machine learning engineer in Lalitpur, Nepal, applying for PhD programs in computational neuroscience. I build radiology systems, have completed an EEG sleep-stage study, and document neural networks from autograd to GPT.
85.14% · Distinction · Eight semesters of full tuition funding
B.Sc. CSIT, Tribhuvan University, 2022–2026. Kathford’s Merit (Jehendar) Scholarship: an entrance-based award followed by seven semester merit awards. Scholarship and issued certificate details
What I do
Radiology systems & machine learning
Full-stack and ML engineering at HOPMEDS HEALTH TECH PRIVATE LIMITED: medical-imaging pipelines, PACS workflows, and DICOM integration with Orthanc and OHIF.
Neural networks, built and explained
Completed GPT-from-scratch study: a character-level decoder transformer, with annotated code, handwritten derivations and experiments from autograd onward.
What I've done
Completed sleep-stage research
Studied generalization to unseen people, comparing spectral baselines, neural models and temporal context. Implementation and internal final evaluation are complete; publication remains pending.
From equations to working models
Seven annotated learning notebooks, three interactive explanations and eight concept lessons connect neural-network code to its mathematics and assumptions.
1 Research
How well does a model trained on some people generalize to others? My completed EEG sleep-stage study examines this question through person-separated evaluation and controlled temporal comparisons. I’m interested in neural decoding, brain–computer interfaces and learning from signals that evolve over time.
2 Projects
- Sleep staging · Internal study completed. Frozen development decisions, reserved-person evaluation and analysis of stage and transition errors. Publication and deployment are pending.
- GPT from scratch · Completed learning project. A character-level decoder transformer exploring the ideas behind GPT-2, following Karpathy’s teaching sequence with my own explanations.
- Motor imagery · Planned. A future EEG decoding experiment; dataset and target classes remain to be selected.
3 Learn with me
The learning notebook is the fullest record of how I work through an idea: code beside explanations, handwritten derivations, saved experiments and recall questions. Start with autograd, follow the makemore series, or explore decoder transformers.
For a shorter visit, try a worked interactive model. Follow gradients through a small network, vary the current driving a leaky neuron, or separate frequency components in a synthetic EEG-like signal.
4 Notes
Six focused notebook-based notes develop ideas from autograd and makemore, including shared-node gradients, conditional probabilities and batch normalization. An EEG essay explains what frequency bands can—and cannot—tell us.
5 Now
Building radiology software and ML systems, and maintaining the learning collection after completing GPT from scratch. The sleep-stage study’s internal evaluation is complete; publication and external-cohort validation are separate next steps. Spiking neural networks remain an active learning interest.
6 CV & Contact
I’m looking for PhD research at the intersection of neuroscience and machine learning. I’d welcome a conversation about neural decoding, temporal models, or a potential fit with your group.