Data Scientist · Machine Learning Engineer · Applied AI
Data to models. Models to decisions.
I connect data quality, model design, evaluation, and application layers to turn complex signals into grounded answers, defensible predictions, and decision-ready intelligence.
End-to-end deliveryData → models → systems → working decisions
Built with evidence, evaluation, and clear handoffs.
Data
Models
Systems
Decisions
74distinct public AI/ML implementations
7+years across engineering & data
12curated public repository collections
01
Flagship agentic RAG
Four systems. Four consequential questions.
A dedicated portfolio of retrieval, tools, graphs, verification, and human control—built for industrial reliability, software engineering, financial research, and portfolio intelligence.
01RetrieveHybrid evidence, not prompt-only answers
02ReasonBounded tools and domain-specific routes
03VerifyCitations, numbers, tests, and provenance
04DecideExplicit limits and human review
01
Industrial reliability
Deployed
ReliabilityOps
Evidence-Grounded Root-Cause Investigation
Business need
Reliability investigations must connect operating conditions, historical incidents, procedures, and statistics before corrective action is considered.
System objective
Retrieve and analyze relevant evidence, rank root-cause hypotheses, verify every material claim, and keep final action behind explicit human approval.
This is research engineering—not investment advice—and material claims must carry supplied evidence IDs.
04
Portfolio intelligence
Deployed
Portfolio RAG
Source-Cited AI Portfolio Assistant
Business need
Recruiters and technical reviewers need a faster way to navigate dozens of repositories without losing the evidence behind each answer.
System objective
Search public portfolio documentation semantically, select relevant project evidence, and return transparent answers with sources, scores, and latency.
Only public portfolio material is indexed; low-support answers expose evidence and limitations.
02
Selected engineering work
Systems built around real technical decisions.
Four distinct non-RAG systems selected for recruiter value: retrieval quality, safe language generation, evidence-backed reporting, and time-series anomaly detection.
01 / Information RetrievalDeployed
Cross-Encoder + Bi-Encoder Ranking
Problem
Balance the speed of dense retrieval with the precision of pairwise relevance scoring.
Engineering decision
Combined a MiniLM bi-encoder with an MS MARCO cross-encoder and benchmarked the two-stage design against lexical baselines on BEIR datasets.
Evidence
Reranking improved MRR@10 by 8.1% on SciFact and 10.7% on NFCorpus, with the latency tradeoff reported explicitly.
Sentence TransformersPyTorchBEIRTransformers.jsHugging Face
Translate business questions into SQL while preventing unsafe or invalid database operations.
Engineering decision
Fine-tuned CodeT5+ with LoRA over a verified multi-schema corpus, then added schema linking, validation, conservative repair, and read-only SQLite execution.
Evidence
Evaluated on 1,040 held-out examples with 56.92% execution accuracy; unsafe queries are blocked before execution.
Turn KPI tables into concise executive narratives without allowing a language model to invent numerical claims.
Engineering decision
Used FLAN-T5 to generate a structured claim plan, then deterministically hydrated values and blocked unsupported claims through a verifier-gated output path.
Evidence
Achieved 100% final displayed claim support across the 250-example evaluation after verifier gating—not a claim of raw-model factuality.
Detect abnormal equipment behavior from multivariate sensor windows while minimizing missed failure events.
Engineering decision
Built an LSTM autoencoder over sensor sequences with reconstruction scoring, threshold selection, baseline comparison, and portable inference artifacts.
Evidence
Reached 0.887 test ROC-AUC and 0.858 failure recall across 918 held-out test windows.
This interactive section shows my full workflow in seven steps: find the evidence, verify it, analyze it, build and test the model, integrate the result, and support a real decision. Select a step to see the tools and proof behind it.
01 / Selected layerevidence → decision
Source
What evidence exists?
Start with the operating signal: structured records, documents, time-series measurements, images, or public datasets.
Methods & tools
SQLDocumentsSensorsAPIsPublic datasets
Evidence in practice
Quality records
Longitudinal sensor data
Research and portfolio corpora
04
Applied in operating contexts
Professional work changes the constraint set.
Industry and research systems must account for confidentiality, uneven error costs, reproducibility, and the people using the output.
Context 01
Hach Company
Quality Data Scientist · 2024 — Present
Turn complex quality information into reliable, inspectable decision support.
≈50potential internal users
2015+source-history coverage
10–15stypical supported answers
Public-safe system architectureHow evidence moves into action
Helps quality teams investigate patterns consistently while preserving evidence, limitations, and confidentiality.
Representative system workstreams
Named at a public-safe level so the engineering logic is visible without exposing protected implementation details.
Move from high-volume telemetry to evidence-led investigation, prioritization, and restoration.
01Network telemetry
02Signal correlation
03Incident context
04Root-cause investigation
05KPI visibility
06Escalation & restoration
02 / repeatable service analyticsOperational reporting automation
Standardize recurring preparation and KPI checks before information reaches operations teams.
01Recurring metrics
02Python / SQL
03Quality checks
04KPI calculations
05Tableau reporting
06Service decisions
01
Network telemetry
Performance metrics, alarms, incidents, equipment, and service events
02
Data preparation
Python and SQL extraction, validation, correlation, and recurring automation
03
Fault context
Device, incident history, time patterns, and operational impact
04
Analytical layer
Anomaly investigation, root-cause analysis, and predictive POC
05
Evidence review
Historical behavior, operational outcomes, and model evaluation
06
KPI visibility
Tableau dashboards for health, incidents, and equipment availability
07
Operational action
Prioritization, escalation, carrier coordination, and restoration feedback
Operating constraints
Large daily signal volumes, recurring incidents, service-impact risk, and multiple Cisco and Juniper environments.
Validation discipline
Reviewed historical incident patterns, operational outcomes, dashboard consistency, and an internal predictive proof of concept.
Public evidence, private boundaries. Professional credibility does not require exposing protected data or internal architecture.
05
Reliability checklist
What makes an ML solution reliable in practice.
Training a model is only step three. This interactive checklist shows the data, evaluation, integration, and user-facing gates required before a model can be trusted in a real workflow.
Active layer · 03
Model
Choose the simplest architecture that improves the operating metric.
Gate: Candidate trained
Leakage
Can the split reproduce the real decision boundary?
Imbalance
Which classes or cases absorb the errors?
Confidence
When should the system abstain or route to review?
Reproducibility
Can the same artifact produce the measured behavior?
Consumption
Does the output expose evidence and limitations?
06
Interactive engineering atlas
Breadth, organized as technical progression.
66 additional public implementations remain after removing every project already featured above. Select a collection, then choose a project to inspect its business need, approach, architecture, skills, live demo, and code.
01AnalyticsEvidence & inference
02Machine learningPrediction & validation
03Deep learningLearned representations
04Language & visionSequence + multimodal systems
05Applied AIGoverned decision systems
Stage 08 · Transformers → RAG & multimodal AI
Transformer & Multimodal AI
Eight additional systems covering long-document QA, instruction tuning, semantic search, vision-language learning, and transformer applications.
The path is deliberate: network operations built systems discipline, business analytics added decision context, research deepened model rigor, and industry work brought those layers together.
01
Sep 2024 — Present
Quality Data Scientist
Hach CompanyUnited States
Applied AI, predictive NLP, quality intelligence, and analytics automation.
Design multi-stage classification workflows across complex quality taxonomies.
Develop retrieval-grounded knowledge workflows and evaluation practices.
Improved latest internal Defect Symptom macro F1 from 56.1% to 57.7% and high-confidence coverage from 38.4% to 46.0%.
Automate recurring analysis and translate model evidence into decision-ready reporting.
PythonSQL ServerNLPTransformersLightGBMPower BI
02
Sep 2022 — Aug 2023
Student Assistant Manager
University of Arizona BookStoresTucson, Arizona
Retail analytics, inventory planning, dashboarding, and team leadership alongside graduate study.
Analyzed 80,000–100,000 monthly sales records across 30+ categories using Excel and Tableau.
Supported inventory decisions associated with an approximately 15% reduction in overstock and shortage conditions.
Earned promotion within four months and trained 8–10 team members on operations, reporting, and data tracking.
