Machine learning and agentic systems. I build the whole thing — the model, the measurement, and the written-down limits.
Lalitpur, Nepal · Kantipur Engineering College · LinkedIn
- Verification systems — speaker and face pipelines where every decision is inspectable: GMM-UBM, LBPH, calibrated thresholds, speaker-clustered confidence intervals.
- Agentic AI — autonomous agent harnesses, tool-use loops, and the engineering discipline that keeps them reliable.
- Evaluation — reproducible splits, honest baselines, and results that survive their own error bars.
Face authentication built on LBPH — no neural embedding anywhere in the identity decision. Active head-turn liveness, per-user conservative calibration, DPAPI-encrypted biometric storage, 15 test modules, and CI on every push.
5.83% EER on ORL at 7 enrolment images · 94.17% TAR at FAR = 1% ·
three-frame median pooling takes LFW from 26.83% to 18.74% EER, reported
with identity-clustered 95% intervals.
Text-independent speaker verification with a GMM-UBM, MAP adaptation, and T-norm.
0.98% EER (95% CI 0.58–1.26%) over 13,315 trials from 50 speakers held out
of UBM training · minDCF 0.219 · every ablation reproducible from committed
score artifacts.
The repository also documents a withdrawn 5.42% figure and the normalisation
cohort bug that produced it, rather than quietly deleting the number.
Ongoing data-structures and algorithms practice, synced from NeetCode — arrays, two pointers, sliding window, stacks, binary search, trees, heaps, graphs.
Python · NumPy · SciPy · scikit-learn · OpenCV · C++ ·
TypeScript / Next.js · pytest · GitHub Actions