Skip to content
NONikhil Obuleni
AI Engineer & Data ScientistCurrently building AI & research systems at CGMHE
AI Research EngineerLLMs / RAGM.S. Data Science / GWWashington, DC

AI Engineer building production-grade AI systems.

I build LLM and retrieval systems that run in production. At GW's Center for Global Mental Health Equity I work on agent pipelines that score clinical transcripts across 25 countries, RAG grounded in clinical rubrics, and the Airflow and AWS pipelines underneath them. Before that, NLP and forecasting at Data Science for Sustainable Development and computer vision at Asteria Aerospace.

View My WorkGitHubLinkedInDownload Resume
Nikhil Obuleni, AI Engineer
Washington, DC
Selected impact
0+
Clinical transcripts scored

Multilingual competency scoring across 25 countries, replacing a fully manual WHO/UNICEF review process.

0+
Patient records in production

Airflow and AWS pipelines into PostgreSQL across 25+ countries, on a $2.9M NIMH grant.

0%
Lower factual error rate

Production RAG grounding LLM output in EQUIP/ENACT clinical rubrics, on held-out multilingual benchmarks.

0%
Weighted F1, disaster forecasting

ARIMA, LSTM and XGBoost ensemble over 200K humanitarian records across 18 countries.

01· About

What I work on

I work on the engineering side of applied AI: agent pipelines, retrieval systems, evaluation, and the data infrastructure that feeds them. Most of what I build runs against real clinical and humanitarian data, so accuracy and traceability matter more than benchmark scores.

At GW's Center for Global Mental Health Equity I build LLM systems for WHO and UNICEF evaluation programs. That covers automating counselor competency scoring that used to be done by hand, grounding model output in clinical rubrics, and running the evaluations that show whether results hold across 14 languages.

Before that I built NLP and LLM-assisted extraction pipelines for NGO reporting at Data Science for Sustainable Development, forecasting models on humanitarian records across 18 countries, and computer vision models for drone systems at Asteria Aerospace.

Now
AI Research Engineer, CGMHE
Based in
Washington, DC
Program
M.S. Data Science, GWU, GPA 3.75
Working on
LLM evaluation, RAG, agent systems
02· Projects

Selected projects

RAG infrastructure, LLM evaluation, generative models, and applied ML. Each card shows the architecture and the measured result.

Filter
medrag-toolkit / hybrid retrieval · citation gate
Dense + sparse · rerank · grounding
PubMedmetformin renal dosing0.73
RxNormmetformin HCl 500mg0.68
PubMedcontrast-induced AKI0.71
OpenFDAlabel: lactic acidosis0.59
PubMedeGFR thresholds review0.54
citation gate3 passages cited

Drag the mix · BM25 over Qdrant at 0%, FAISS at 100%

01RAGLLM Infrastructure

medrag-toolkit modular RAG for medical QA

A modular RAG framework for medical question answering with pre-built connectors to PubMed (36M+ abstracts), RxNorm, and OpenFDA. Every response enforces citation grounding and passes an automated hallucination-detection layer. Hybrid retrieval combines dense FAISS with sparse BM25 over Qdrant, and a streaming FastAPI layer supports local backends (Ollama, vLLM) as well as cloud LLMs.

Sources
PubMed, RxNorm, OpenFDA
Retrieval
Dense + sparse hybrid
Benchmarks
MedQA, MedMCQA
LangChainLlamaIndexFAISSQdrantvLLMOllamaFastAPIDocker
View on GitHub
clinical-rag / 5,000+ WHO · NIMH guidelines
Request trace · RAGAS result
  1. query
  2. guardrail
  3. retrieve
  4. rerank
  5. generate
  6. score
response relevance--
hallucination reduction--
verification time saved--

Replaying a traced request · benchmarked with RAGAS vs. GPT-4

02RAGEvaluation

Clinical RAG Evaluation System retrieval over clinical guidelines

A RAG pipeline over 5,000+ WHO and NIMH clinical guidelines using semantic chunking, deployed on AWS Lambda behind FastAPI. LangChain and Pinecone handle retrieval, prompt-injection guardrails sit on the input path, and evaluation logging runs through LangSmith. Benchmarked with RAGAS against GPT-4, then tested with 3 clinical reviewers.

Response relevance
91%
Hallucination reduction
38%
Verification time
45% faster
LangChainPineconeGPT-4FastAPIAWS LambdaRAGASLangSmith
Repo not yet public
planmoe / room-constraint graph → MoE
Gating network · expert routing
GCN x3gatingroutes to

layout

Auto-cycling · hover an expert · INT8, single GPU

03Generative AIDeep Learning

PlanMoE generative floor plan designer

A custom Mixture of Experts with a 3-layer GCN encoder and a transposed-convolution MaskDecoder. Each expert specializes in one room-type layout, and a learned gating network routes room-constraint graphs to the right expert at inference. Deployed as a HuggingFace Space with a FastAPI backend, with INT8 post-training quantization bringing inference inside a real-time budget.

Inference latency
52% lower
Serving
Single GPU, real time
Routing
Learned gating network
PyTorchMixture of ExpertsGCNHouseGAN++FastAPIHuggingFace Spaces
View on GitHub
mh-predict / multi-modal ensemble
Free text + records → 3 categories
sleepappetiteworryenergyfocusmood

TF-IDF features

ensemble
cat 1cat 2cat 3

Cycling a screening record · TF-IDF + ensemble over 10K+ records

04NLPHealthcare

Mental Health AI Prediction System multi-modal NLP and ML

A multi-modal NLP and ML pipeline predicting across 3 diagnostic categories, trained on 10K+ clinical screening records with TF-IDF feature extraction and ensemble classification. Paired with an emotionally-aware chatbot built on TensorFlow and sequence-to-sequence modeling, informed by clinical ground truth from GWU's mental health research program.

Accuracy
94%
Diagnostic categories
3
Training records
10K+
PythonTensorFlowNLPscikit-learnFlask
Repo not yet public
streaming-rec / 10M+ viewer records
Graph neighbourhood · churn signal
Viewer AThrillerDocsViewer BK-dramaViewer CComedy
12
Churn p
Viewer A

Auto-cycling · hover a node to inspect · 2 linked nodes

05Deep LearningRecommenders

Content Recommendation & churn prediction

A graph neural network and collaborative filtering recommendation engine over 10M+ viewer records, with an XGBoost and LSTM/ANN churn model built on the same subscriber features. GNN embeddings over the viewer-title interaction graph feed the recommender, so retention strategy and recommendation stay consistent.

Engagement lift
35%
Churn accuracy
92%
Churn reduction
20%
PythonPyTorchGNNXGBoostLSTM
View on GitHub
Also on GitHub

More experiments and code on GitHub.

03· Experience

Experience

LLM systems for global mental health research, ML and data engineering for humanitarian reporting, and computer vision for drone platforms.

CurrentMar 2026 to Present· Washington, DC

AI Research Engineer

Center for Global Mental Health Equity (CGMHE)

LLM agent pipelines, retrieval systems, and evaluation for WHO and UNICEF mental health programs, plus the data platform they run on.

  • Shipped an LLM agent pipeline (LangChain, OpenAI API) that automates counselor competency scoring across 26K+ multilingual clinical transcripts in 25 countries, cutting per-session review time from 45 minutes to 4 and replacing a fully manual WHO/UNICEF evaluation process.
  • Designed a production RAG system (FAISS, LangChain, FastAPI) grounding LLM output in EQUIP/ENACT clinical rubrics through chunking, reranking, and a hallucination-flagging layer, reducing factual error rate by roughly 40% on held-out multilingual benchmarks.
  • Ran LLM evaluation experiments (RAGAS, DeepEval) comparing GPT-4 and Claude against human-annotated ground truth across 14 languages, identifying a 23% calibration gap on low-resource languages that fed into a joint OpenAI collaboration on clinical LLM deployment.
  • Built Airflow and AWS (S3, EC2) pipelines ingesting 500K+ patient records into PostgreSQL across 25+ countries on a $2.9M NIMH grant, cutting integration latency 83% and reporting time 60%.
LangChainOpenAI APIClaude APIFAISSFastAPIRAGASAirflowAWS
01 / 03
04· Skills

Skills

Grouped by what each tool is for. Everything listed here shows up in the work above.

58 / 58 matching

AI / Machine Learning

13

Model building, from classical ML through deep learning and computer vision.

PythonPyTorchTensorFlowscikit-learnXGBoostLightGBMBERTspaCyHugging FaceGNNLSTMYOLOv5Faster R-CNN

Generative AI

9

LLM application work: agents, retrieval, and the prompting layer around them.

OpenAI APIAnthropic (Claude) APILangChainLangGraphRAGEmbeddingsVector SearchPrompt EngineeringMulti-Agent Systems

RAG Infrastructure & Evaluation

10

Vector stores, serving, and the tooling that shows whether a system actually works.

FAISSPineconeQdrantpgvectorvLLMOllamaRAGASDeepEvalLangSmithMLflow

Data Engineering

9

Pipelines and storage, so models run on a schedule instead of on my laptop.

SQLPostgreSQLMongoDBApache AirflowBigQueryPySparkPandasETL PipelinesREST APIs

Cloud / Engineering

7

Deployment, containers, and CI.

AWS (Lambda, EC2, S3, SageMaker)GCPDockerFastAPIGitHub ActionsGitCI/CD

Analytics / Visualization

10

Statistical methods and the reporting layer stakeholders actually read.

Statistical ModelingTime Series ForecastingAnomaly DetectionSHAP ExplainabilityA/B TestingCausal InferenceTableauPower BIMatplotlibPlotly
05· Education

Education

Coursework behind the applied work, and the fellowship that funded it.

M.S. Data Science

Jan 2025 to Dec 2026

George Washington University

GPA 3.75. Global Leaders (CCAS) Fellowship, $17,100 (2025). Washington, DC.

Coursework
Machine LearningDeep LearningNatural Language ProcessingStatistical ModelingData MiningData VisualizationDatabase ManagementGIS
06· Contact

Let's build something useful.

I am open to AI and ML engineering roles, and happy to talk about retrieval, LLM evaluation, or anything that has to run reliably against real-world data.

Email
nikhil.obuleni@gwu.edu
Send an email

Washington, DC · (571) 478-2007

GitHubnik-hill-323↗LinkedInnikhil-obuleni↗ResumePDF, one page↗