I keep busy teams organized, communicate clearly with people at every level, and dig into data or technical problems when the work calls for it. Open to roles in research, HR, operations, and admin. US citizen, no sponsorship needed.
I'm someone who likes keeping systems and people in sync. That might mean running the front desk of a busy tax office, sourcing candidates for an HR product, or teaching myself a technical project from scratch. The through-line is the same: listen carefully, stay organized, and finish what I start.
Most recently I worked as an HR Product Intern at EMRA AI and as Office Manager at H&R Block, where I was promoted within six weeks. On my own time I also built an EEG decoding project that classifies imagined movement with up to 83% accuracy. I like work where communication, detail, and curiosity all matter.
I'm open to roles across research, HR, operations, and administration. US citizen, authorized to work with no sponsorship needed.
Experience & Education
Client-facing leadership, product support, a Power BI project, certifications, and self-directed technical work.
Work
Independent Researcher
Self-directed EEG and BCI Research
June 2026 to Present
Built a four-class motor imagery decoder (EEGNet, 3,508 params) on BCI Competition IV-2a for left hand, right hand, feet, or tongue from four seconds of scalp EEG. Cross-day evaluation: best subject 83%, nine-subject mean 57.7%.
83% best-subject accuracy (chance 25%)
EEGNet · 3,508 parameters · braindecode
Cross-day train/test · 9 subjects
HR Product Intern
EMRA AI
April 2026 to June 2026 · Remote
Supported an HR hiring product by sourcing candidates, coordinating interview schedules, screening applicants, and relaying client feedback to guide product improvements.
Candidate sourcing & screening
Interview & meeting coordination
Client feedback for product
Office Manager
H&R Block
December 2023 to April 2026 · Queens, NY
Promoted from Receptionist within six weeks. Managed the front desk for a seven-person tax office, scheduled six preparers each season, and handled high-volume calls and sensitive financial information with care.
1,500+ appointments per season
80 to 150+ calls/day at peak
Promoted within six weeks
Confidential ID & IRS PIN handling
Front Desk Associate
Elite Fitness
September 2023 to April 2024 · New York
Welcomed members and ran the front desk at a private gym of 30 to 40 members. Handled memberships, questions, and kept the desk and records in order.
Member welcome & front desk
5 to 10 new memberships/month
Records & desk upkeep
Education
Associate of Science, Business Administration
LaGuardia Community College (CUNY), New York, NY
In progress
Expected 2027
Progress40%
Part-time study alongside full-time seasonal work. Building a foundation in operations, management, and professional communication.
Focus areas
Business operations
Organizational management
Professional communication
Records & administration
Self-Directed Study: EEG, BCI & Applied ML
Independent
In progress
2025 to Present
Progress70%
Structured self-study behind the BCI Competition IV-2a decoder: literature, experiment workflow, compact CNNs for EEG, and clear reporting.
Focus areas
EEG preprocessing (4-38 Hz)
Four-class motor imagery
EEGNet / braindecode
Cross-day evaluation
Projects
HR Attrition Dashboard
Power BI
2025
Built an interactive Power BI dashboard to study employee attrition and guide HR retention decisions.
Imported raw CSV data and cleaned it in Power Query by removing duplicates, fixing errors, and setting data types
Created an AttritionCount column and DAX measures to calculate the attrition rate as a percentage
Designed bar charts, KPIs, pie charts, and tables to present clear and useful insights
Certifications & Awards
Project Management
Google / Coursera
2025
Agile with Jira
Atlassian / Coursera
2025
Featured project
Reading movement intent from EEG
Decoding imagined movement from scalp EEG. 83% accuracy on a four-way choice where guessing scores 25%.
I trained a compact neural network (EEGNet, 3,508 parameters) to tell which of four movements a person was imagining: left hand, right hand, feet, or tongue, from four seconds of scalp EEG. On held-out recordings from a different day, it reaches 83% on the best subject and 57.7% averaged across nine people. Chance is 25%.
83%
Best-subject four-class accuracy (chance 25%)
57.7%
Mean across nine subjects (38% to 84%)
3,508
Parameters. Trains on a laptop CPU
61.7%
Pooled model beats nine personal models
Brain-computer interfaces let people control devices by imagining movement. The hard part is that EEG is noisy, low-resolution, and different for every person. I built a decoder for the four-class motor imagery task on the BCI Competition IV-2a dataset: given four seconds of scalp EEG, decide which of four movements the person was imagining.
The model is EEGNet, deliberately tiny at 3,508 parameters, since EEG datasets are small and large models just memorise. I band-pass filtered to 4-38 Hz (the mu and beta rhythms where motor imagery lives), cut the recordings into labelled trials, and trained on one session while testing on a second session recorded on a different day, so nothing leaks between them.
The best subject reaches 83% against a 25% chance baseline. Across all nine subjects the average is 57.7%, ranging from 38% to 84%. That 45-point spread reflects a real, unsolved problem in the field, not a flaw in the pipeline. Training one shared model on all nine subjects instead beats the per-subject average, reaching 61.7%.
Key results
83.0% four-class accuracy on the best subject (chance: 25%)
78.1% mean over five retrains, range 70.5-81.6% (the honest spread)
57.7% mean across all nine subjects, range 38.5-84.0%
61.7% for a single model pooled across all subjects, beating nine personal models
3,508 parameters. Runs on a laptop CPU in minutes
Tested on a separate recording session from a different day
Results figures
Tap any figure to enlarge.
Methods
Data: BNCI2014_001 (BCI Competition IV-2a), 9 subjects, 26 channels at 250 Hz, two sessions each. Preprocessing: 4-38 Hz band-pass, rescaled to microvolts, epoched into 4-second trials around each cue (288 train / 288 test per subject). Model: EEGNet via braindecode, Adam at lr 0.0625, batch size 32, 100 epochs. Evaluation: session 1 trains, session 2 tests. A cross-day split, never a random shuffle of trials.
Moments
Gallery
A few frames from work, demos, and everyday curiosity.
Skills & Interests
Technical
Microsoft Office (Outlook, Word, Excel, PowerPoint)Power BIJiraPower QueryDAXAMP scheduling portalPython
Administrative
Calendar and meeting coordinationScheduling for multiple team membersHigh-volume phone handlingFront desk and reception coverageFiling and recordsData entrySensitive document handling
Attention to detailConfidentiality and discretionProblem solvingDependabilityTime management under deadlines
Languages
English (fluent)Tibetan (native)
Contact
Open to roles in research, HR, operations, administration, and related work where clear communication and follow-through matter. Email or LinkedIn works best.
