Skip to content
View nik-hill-323's full-sized avatar

Block or report nik-hill-323

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
nik-hill-323/README.md
Nikhil Obuleni. AI research engineer. Retrieve, ground, evaluate, decide.

I build LLM systems for people who cannot afford a wrong answer: clinicians, counselors, and field teams in low-resource settings. Most of my code falls into one of four stages, and I keep a repo for each.

retrieve & ground model evaluate decide

medrag-toolkit Medical question answering over PubMed, OpenFDA and RxNorm. Hybrid dense + BM25 retrieval. Every sentence is checked against its source; anything unsupported is dropped, and it abstains rather than guesses.

PlanMoE Mixture of experts over room-constraint graphs, INT8 quantized, on HF Spaces.

sepsis-prediction BiLSTM that flags ICU sepsis risk 6 to 12 hours early.

skiavox Chest X-ray ensemble with Grad-CAM and DICOM in.

mindscope Mental health NLP benchmark, 15+ conditions, BERT and Claude baselines.

ragcheck Scores retrieval, faithfulness, relevance and citation coverage separately, and prints the exact sentences a RAG answer made up. Offline judge for CI, LLM judge for depth, a gate command that fails the build.

abtest-kit Power analysis, z and Welch tests, Bayesian expected loss, always-valid sequential tests, CUPED. Each method is tested against an independent reference. The peeking demo: naive t-test 32% false positives, mSPRT 1.8%.

log

when what
Sep 2026 Published ragcheck and abtest-kit. Both green on CI across Python 3.10 to 3.12.
Aug 2026 medrag-toolkit and PlanMoE went public. Rebuilt the portfolio site on Next.js.
Mar 2026 Joined the Center for Global Mental Health Equity at GWU as an AI research engineer. The agent pipeline I shipped scores counselor competency across 26K+ transcripts in 25 countries and cut per-session review from 45 minutes to 4.
2025 – 2026 ML engineer at Data Science for Sustainable Development: NLP extraction over sustainability reports and ARIMA/LSTM/XGBoost forecasting on 200K+ humanitarian records across 18 countries.
Jan 2025 Started the M.S. in Data Science at George Washington University. GPA 3.75, finishing Dec 2026.
2023 Drone object detection and predictive maintenance at Asteria Aerospace, Bangalore.

what the evaluation work found

Comparing GPT-4 and Claude against human-annotated ground truth across 14 languages turned up a 23% calibration gap on low-resource languages. A RAG layer grounded in WHO EQUIP/ENACT rubrics, with reranking and hallucination flagging, brought the factual error rate down by about 40% on held-out multilingual benchmarks. Those findings feed a joint study with OpenAI on clinical LLM deployment. ragcheck is the open-source distillation of that measurement work.

toolbox

Python, R, PyTorch, TensorFlow, scikit-learn, FastAPI, Docker, AWS, GCP, PostgreSQL, MongoDB, GitHub Actions, git, Linux

Plus the parts without icons: LangChain and LangGraph, the OpenAI and Anthropic APIs, FAISS, Pinecone, Qdrant and pgvector, RAGAS, DeepEval and LangSmith, Hugging Face, spaCy, XGBoost and LightGBM, PySpark, Airflow, BigQuery, MLflow, Streamlit.

activity

Contributions, commits, repos, languages and a 26-week sparkline, regenerated daily

These cards are rendered by scripts/build_stats.py in this repo, refreshed every morning by a GitHub Action. No third-party widget service.

elsewhere

portfolio · resume · linkedin · nikhil.obuleni@gwu.edu

Open to AI engineer and data scientist roles from 2027. If you work on AI for health and want a second pair of eyes on an evaluation setup, my inbox is open.

Popular repositories Loading

  1. nik-hill-323 nik-hill-323 Public

    Profile README

    Python

  2. cs50 cs50 Public

    Forked from endiliey/cs50

    📗 Problem sets for CS50 course by Harvard University

    C

  3. Exploratory-Data-Analysis-of-WMATA Exploratory-Data-Analysis-of-WMATA Public

    Exploratory analysis of WMATA Metrorail and Metrobus ridership: boardings by route, time of day and day of week, plus origin-destination patterns. R.

    HTML

  4. Data_mining_Final_project Data_mining_Final_project Public

    Data mining course project: cleaning, demographic and geographic EDA, and modelling in Python.

    Jupyter Notebook 1

  5. Analysis-of-HR-Attrition-Patterns Analysis-of-HR-Attrition-Patterns Public

    Statistical and ML analysis of IBM employee attrition: drivers of turnover and models that flag employees at risk of leaving. R.

    HTML

  6. Nikhil_Obuleni Nikhil_Obuleni Public

    Personal portfolio site. Next.js 15 static export, deployed to GitHub Pages.

    TypeScript