I architect and deploy production AI evaluation systems, large-scale randomized experimentation frameworks, and interpretable machine learning pipelines. My work bridges the gap between deep algorithmic rigor (dense vector retrieval, proposition entailment, knowledge graphs) and tangible enterprise unit economics.
| Key Performance Indicator | Benchmark / Outcome | Engineering Mechanism |
|---|---|---|
| Operational Deflection | 44.5% autonomous resolution ($20,760/mo savings) | Calibrated RQI decision gate & human fallback |
| Hallucination Interception | 100% out-of-scope query containment | Reference-free RAG Triad entailment scoring |
| Pedagogical Lift | +14.9% post-remediation student score increase | Directed Acyclic Graph (DAG) prerequisite gap tracing |
| Experimentation Scale | 4,000 learners across 4 randomized arms | Bonferroni-Holm family-wise error control & HTE |
| Lead Funnel Precision | Priority scoring on 3,000 prospective leads | SMOTE oversampling + F2-calibrated XGBoost & SHAP |
flowchart LR
A["Raw Queries & Events"] --> B["Dense Vector Retrieval<br/>(ChromaDB / MiniLM)"]
B --> C{"RAG Triad & RQI Gate<br/>(Threshold Cutoff = 0.72)"}
C -- "High Quality (>= 0.72)" --> D["Autonomous Resolution<br/>(44.5% Deflection)"]
C -- "Uncertain / Fringe (< 0.72)" --> E["Expert Human Queue<br/>(Zero Hallucinations)"]
D --> F["Telemetry & Unit Economics<br/>($20,760 Monthly Savings)"]
Stack:
Python·SentenceTransformers·ChromaDB·Streamlit·In-Flight Telemetry
- Packaged enterprise dual-tab application combining live student doubt simulation with real-time financial telemetry.
- Translates AI operational reliability into $20,760/month projected cost savings against human tutoring baselines.
Stack:
SentenceTransformers·Proposition Entailment·Scikit-learn·Streamlit
- Reference-free audit system for unlabelled student queries evaluating Context Relevance, Faithfulness, and Answer Relevancy.
- Established an empirical RQI cutoff of 0.72, intercepting 100% of out-of-scope hallucination baits.
Stack:
ChromaDB·Vector Retrieval·Cosine Calibration·Streamlit Dashboard
- Dense vector retrieval pipeline over 500 curriculum chunks with confidence-calibrated human fallback routing.
- Achieved >85% routing accuracy, safeguarding learning integrity with automated escalation of low-confidence queries.
Stack:
XGBoost·Random Forest·SMOTE·SHAP·Google Sheets Scorecard
- Lead scoring engine correcting extreme class imbalance (~18% prevalence) to rank 3,000 prospective learners.
- Calibrated decision thresholds on Precision-Recall curves maximizing F2-score to minimize missed student conversions.
Stack:
SciPy·Bonferroni-Holm Correction·Welch t-tests·Heterogeneous Treatment Effects
- 4-arm randomized controlled trial evaluating micro-videos, quizzes, and long-form lectures across 4,000 learners.
- Identified that interactive micro-videos neutralize the mobile device friction gap (80.2% mobile completion).
Stack:
NetworkX DAG·Cognitive Modeling·Multi-Corpus Ingestion·Dual-Tab Streamlit
- Traces exam errors through a Directed Acyclic Concept Dependency Graph back to root-cause prerequisite failures.
- Drives an empirical +14.9% test score lift by replacing generic answers with targeted remediation assets.
Click to expand full repository catalog (Ordered newest to oldest)
| Repository | Domain | Year | Focus |
|---|---|---|---|
| stakeholder-rag-deployment | Production AI & Telemetry | 2026 | Financial ROI telemetry & live RAG deployment |
| reference-free-rag-triad | AI Evaluation & Safety | 2026 | Reference-free hallucination audit pipeline |
| multisource-diagnostic-remediation | Cognitive Systems & Graphs | 2026 | Directed acyclic concept prerequisite remediation |
| content-optimization-ab-testing | Causal Inference & Stats | 2026 | 4-arm randomized controlled trial with HTE |
| enrollment-propensity-model | Predictive ML & Attribution | 2026 | Interpretable lead propensity scoring with SHAP |
| rag-doubt-resolution | Vector Search & IR | 2026 | ChromaDB dense semantic retrieval with fallback |
| renewable-energy-orchestrator | Multi-Agent Systems | 2026 | Agentic renewable grid energy orchestration |
| khan-academy-data-analysis | Educational Analytics | 2026 | In-memory SQLite reading fluency ETL pipeline |
| LLM-Augmented-Enrollment-Intelligence-System | NLP & Lead Analytics | 2026 | Sentiment analysis and LLM-assisted lead scoring |
| data-science-capstone | Foundational Applied DS | 2025 | IBM Professional Certificate SpaceX Falcon 9 analytics |
| DSEcosystemsnotebook | Foundational Data Science | 2024 | IBM Data Science methodology & ecosystem lab |
| sjnewrepository | Archive | 2024 | Early exploratory development workspace |
- AI, NLP & Retrieval: SentenceTransformers, ChromaDB, Semantic Embedding Dot-Product, Proposition Entailment, RAG Triad.
- Machine Learning & Modeling: Scikit-learn, XGBoost, Random Forest, Logistic Regression, SMOTE (imbalanced-learn), SHAP, NetworkX.
- Statistical Inference & Causal Analysis: Two-sample Z-tests, Welch's t-test, Bonferroni-Holm multiple testing correction, Heterogeneous Treatment Effects (HTE).
- Engineering, Telemetry & Web: Python, Pandas, NumPy, SQLite3, Streamlit, Matplotlib, Seaborn, Git, Docker, REST APIs.