Graduate TA
Sam Houston State University
Teaching lab and tutoring sessions for stats and data science. Helping students work through R, Python, and statistical modeling: regression, ANOVA, the usual.
Chirag Lamsal
data, stats, and code that ships.
Houston, TX
est. somewhere in Nepal
me, allegedly
Hi! I'm a data analyst and software engineer specializing in statistics and machine learning. Currently an M.S. student at Sam Houston State University.
chapter one
Experience
Sep '25 →
Software Engineer
Sandbox Software
Shipped payment integrations that bumped platform revenue by 29%. Built a real-time auction marketplace (+11% engagement) and ran AWS / Docker / Cloudflare deploys at 99.8% uptime.
+29%
metrics #001
The Poster
STAT #2026
Abstractor · arXiv Classifier
An end-to-end NLP pipeline that classifies a paper’s title and abstract into one of 8 academic domains. 61,640 papers fetched from the arXiv API; 8 classifiers compared on macro-F1.
0.731
macro-F1
61,640
papers
8
classifiers
Model leaderboard
LR-L2 · TF-IDF
0.731
LinearSVC · TF-IDF
0.727
LinearSVC · SVD
0.724
LR-L2 · SVD
0.722
Multinomial NB
0.705
XGBoost · SVD
0.704
Random Forest
0.618
winner → Ridge logistic on TF-IDF, trained in 4.1s.
next time → re-fit the vectorizer on train-only (mild leakage in current run); try SciBERT to squeeze the CS↔EE confusion.
- Preprocessing
- LaTeX normalization (recovers
\betaas beta), scispaCy lemmatization, corpus-based stopwords, and gensim phrase detection for terms like neural_network. - Validation
- Before modeling: ~280K chi-square tests with BH correction, plus PERMANOVA on TF-IDF cosine distances. Pseudo-F = 6.601, p = 0.001 → classes are meaningfully separable.
- Stack
- Pythonscikit-learnscispaCygensimXGBoostStreamlit
file #001
Case File
SUBJECT #2026
The Vitamin D Mystery
An investigation into the association between serum Vitamin D and Diabetes risk, drawing on NHANES survey data.
- Hypothesis
- Serum Vitamin D levels are inversely associated with Diabetes prevalence across demographics.
- Method
- Weighted logistic regression accounting for complex survey design. Confounding adjustment for BMI, age, and physical activity.
- Instruments
- Rggplot2survey pkgANOVA
- Findings
- Evaluated via AUC / ROC with confidence intervals. Visualized factor relationships across subgroups.
✓filed under: health-data · epidemiology · weighted-models
drawing
Blueprint
DRAWING #2024
AssistAI · RAG Pipeline
A retrieval-augmented chatbot that answers questions grounded in a user’s own document knowledge base.
- Retrieval
- Documents chunked & embedded into a vector store (FAISS/Chroma); queries resolved via semantic search.
- Generation
- Retrieved context injected into the prompt so the model answers from source material, not just priors.
- Service
- FastAPI endpoints expose chat + ingest; streaming responses wired to the client.
- Stack
- PythonLangChainFastAPIFAISS
SHEET 1 OF 1SCALE = 1 : 1DRAWN BY: C
the arsenal
Skills
The Lab
where numbers get interrogated
RPythonSASSQLC++StatisticsScikit-learnXGBoost
The Forge
where things get shipped
AWSDockerDjangoPostgreSQLGitREST APIsNginxCelery
say hi