Five technical case studies
AI systems.
Measured, not imagined.
LLM orchestration, model evaluation, generative applications, and statistical learning—documented through architecture, automated tests, benchmarks, and measured results.
01LLM systems engineering2026
Hybrid AI Gateway
A privacy-aware gateway that routes requests between a local model and a cloud model using sensitivity, task type, complexity, and cloud-permission signals.
01Classify request
02Apply privacy gate
03Select model path
04Validate or fall back
106automated tests
7 / 7end-to-end cases
CIpassing
ProblemCloud-only AI can expose sensitive context; local-only AI can struggle with demanding reasoning. The system needed to select the right execution path without making privacy a soft preference.
Technical buildA deterministic routing layer, privacy hard gate, local Gemma 3 execution through Ollama, cloud GPT-OSS execution through GroqCloud, structured outputs, fallback handling, Docker support, and automated evaluation.
Evidence106 automated tests passing, 7/7 end-to-end evaluation cases passing, Docker verified, GitHub Actions CI green, and privacy routing validated.
PythonFastAPIOllamaGemma 3GroqGPT-OSSPytestDockerGitHub Actions
02LLM benchmarking2026
Local vs Cloud AI Lab
A practical benchmark comparing Gemma 3 4B running locally through Ollama with GPT-OSS 120B running through GroqCloud.
01Load JSON cases
02Run local + cloud models
03Grade outputs
04Aggregate results
34model requests
90%local grounded QA
100%cloud grounded QA
ProblemDetermine when a small local model is sufficient and when a larger cloud model is worth using across privacy, cost, internet dependency, output reliability, reasoning, and latency.
Technical buildA V1 benchmark of 34 model requests across instruction following and structured extraction, grounded QA and abstention, and reasoning, with JSON test cases, custom automated graders, cost analysis, and combined result reporting.
EvidenceV1 complete. Gemma 3 4B reached 90% grounded-QA accuracy and 80% reasoning correctness; GPT-OSS 120B reached 100% on both. Median latency was 0.92 seconds locally and 0.48 seconds in the cloud environment tested.
PythonOllamaGemma 3 4BGroqCloudGPT-OSS 120BCSVJSONAutomated graders
03Retrieval-augmented generation2026
Local RAG Evaluation Lab
A privacy-first local RAG system that combines dense vector retrieval and BM25 through Reciprocal Rank Fusion, then evaluates retrieval quality before generation.
01Ingest local documents
02Retrieve vector + BM25
03Fuse rankings with RRF
04Evaluate and generate
1.000Recall@4
0.893hybrid MRR
17evaluation cases
ProblemPrivate documents need useful question answering without being uploaded to a cloud service. The retrieval layer also needs measurable evidence that it is finding the correct document and page—not just producing plausible answers.
Technical buildA local document-ingestion and chunking pipeline, dense vector search, BM25 lexical retrieval, hybrid RRF fusion, grounded generation through Ollama, retrieval diagnostics, automated evaluation, regression gates, and GitHub Actions CI.
EvidenceEvaluated on 17 labeled cases at Top-K=4. Vector and hybrid retrieval both achieved 1.000 document/page hit rate and 1.000 Recall@4; the hybrid pipeline achieved an MRR of 0.893.
PythonLangChainOllamaVector searchBM25RRFAutomated evaluationGitHub Actions
04Generative AI application2026
Pet Thought Generator
An interactive GenAI application that converts pet context and a configurable chaos score into a generated inner monologue.
01Collect pet context
02Apply chaos control
03Construct prompt
04Render model response
Geminimodel API
Streamlitinteractive UI
.envsecret isolation
ProblemTurn a direct model call into a usable product flow with structured user controls, prompt construction, response rendering, and protected credentials.
Technical buildA Streamlit interface, dynamic prompt orchestration, Google Gemini API integration, controllable generation through a chaos score, and environment-based secret management.
EvidenceA working end-to-end LLM application that connects structured UI input to model inference and presents the response in a browser.
PythonStreamlitGoogle Gemini APIPrompt orchestrationEnvironment variables
05Statistical learningGeorgia Tech
Youth Mental Health Analysis
An adjusted logistic-regression study of screen behavior, physical activity, and mental-health risk using the 2023 National YRBS.
01Prepare complete cases
02Define exposure + outcome
03Fit logistic model
04Diagnose + interpret
9,648analysis records
1.1118unadjusted odds ratio
0.0969p-value
ProblemEstimate whether frequent social-media use was associated with mental-health risk after accounting for multiple behavioral and demographic predictors.
Technical buildComplete-case data preparation, exploratory analysis, adjusted logistic regression, odds ratios, confidence intervals, ROC/AUC, confusion matrix, VIF, and calibration assessment.
EvidenceAnalyzed 9,648 records. Mental-health risk was 24.0% among non-frequent social-media users and 32.6% among frequent users; the unadjusted association was not statistically significant (OR 1.1118, 95% CI 0.9811–1.2599, p=0.0969).
PythonStatistical learningLogistic regressionROC/AUCVIFCalibrationData visualization