Work & case studiesEvidence before claims.

Every project,
open to inspection.

Explore the questions, decisions, and artifacts behind our analysis, prototypes, concepts, and engineering methods.

Each entry identifies what was made, its source, and its limitations.

Independent analysis

From a ride record
to a clearer question.

Explore the Cyclistic source figures, model and disclosed assumptions.

Inspect the case study
Cyclistic daily bike ride counts across 2025 and 2026.
Python repository artifact · source run unverified

Work, in motion.

Five perspectives on a technical problem. Follow the work, then inspect the evidence and its boundaries.

Browse the complete index
Applied AI evaluationPublished engineering method
Concept preview · evaluation method
Applied AI & Automation

Production RAG Validation Suite

A published evaluation method for testing retrieval quality, groundedness, and failure boundaries before a knowledge assistant is released.

Hybrid retrievalCross-encoder rerankingGroundedness evaluation
Workflow automationInterface concept
Concept interface · illustrative flow
Systems & Software

Automated Incident Dispatcher

A concept interface for webhook intake, signature checks, and routing an incident to an on-call owner.

Interface designWebhook flowRouting contracts
Data engineering & BIIndependent portfolio project
Python repository artifact
Source figurePython analysis artifact
Python chart of daily bike ride counts, with date on the horizontal axis and ride count on the vertical axis.

Daily ride counts · Python analysis artifact. The chart’s source run is separate from the BI demonstration export.

Data Analytics & BI

Cyclistic Ridership Analysis

Python analysis artifacts, a dimensional BI model, and a financial scenario explorer.

Python / PandasPower BI / DAXStar Schema
Systems & cloudInteractive system prototype
Illustrative system map
Systems & Software

Cloud Architecture Visualizer

An interactive system map for following data through a knowledge assistant, API synchronization, and a BI pipeline.

TypeScriptReactNext.js
Operational engineeringPublished decision framework
Illustrative decision framework
Systems & Software

Workflow Automation Feasibility Model

A decision framework for choosing between deterministic scripts, queued webhooks, AI assistance, and human review.

Process mappingRisk assessmentMaintenance analysis
Scroll to explore

Projects, prototypes
& published methods.

6 entries
Project 04 / 06
Published engineering methodApplied AI evaluation

Production RAG Validation Suite

A published evaluation method for testing retrieval quality, groundedness, and failure boundaries before a knowledge assistant is released.

  • Hybrid retrieval
  • Cross-encoder reranking
  • Groundedness evaluation

Concept preview · evaluation method

Question, contribution & scope
The question
A convincing demo answer does not establish retrieval quality or reliable handling of unsupported questions.
The approach
Evaluate chunking, hybrid retrieval, reranking, citations, and refusals against a representative question set.

A technical guide with a readiness checklist and a retrieval trade-off table.

Scope: A written methodology, not an implemented benchmark product or a published scored evaluation run.

4 of 6: Production RAG Validation Suite

Trace the system.
Inspect the boundary.

The architecture explorer is a working frontend prototype. Follow a stage to see its inputs, outputs, and failure boundaries.

Interactive concept explorer

Follow the architecture.

Choose a workflow. Inspect its contracts, outputs, and failure paths.

Ground language models in authoritative corporate documents, enforce citation boundaries, and maintain a deterministic human escalation path for low-confidence inference.

Illustrative design / no live system connection
Human Review Escalation

Passages with cross-encoder rerank confidence < 0.78 route to human triage with exact retrieved document context.

Stage 01 / Software Systems

Document Intake

Parse heterogeneous enterprise documents, extract high-fidelity text and tabular structures, and strip non-authoritative boilerplate before indexing.

Contract invariant

Cryptographic SHA-256 hash verified; metadata normalized before parsing.

Failure mitigation

Corrupt or encrypted payloads quarantined to Dead-Letter Queue.

Output artifact

Sanitized markdown AST with tabular preservation

Stage 1 of 4

Have a problem that does not fit a category?

Most real systems cross more than one discipline. Bring the constraint and we will map the smallest reliable path through it.