Paul Martin McNeill's Portfolio

Building Production RAG Pipelines: Lessons from Real Deployments

Wed, 01 Apr 2026 00:00:00 +0000

A demo RAG pipeline takes a weekend. A production one takes months — and most of the cost is in the parts you can’t see in a notebook: drift detection, eval suites, caching strategy, and what happens when retrieval gets it wrong in front of a user.

Chunking is not preprocessing — it’s product design

Most RAG fails because the chunking strategy was treated as a one-line decision…

Retrieval without grounding is just search

If your model can’t cite the chunk it pulled from, you can’t debug a hallucination…

The eval harness is the real codebase

Models change. Prompts change. Embeddings change. Without an automated regression suite you’ll ship silently broken retrieval and never know…

Why I Moved from Jupyter Notebooks to End-to-End ML Pipelines

Sat, 14 Feb 2026 00:00:00 +0000

Notebooks are wonderful for exploration. They are a disaster for deployment. This post is about the moment that distinction stopped being theoretical for me, and what I changed in my workflow as a result…

The hidden cost of “it runs in my notebook”

State that lives only in cell outputs. Cell-execution order that nobody else can reproduce. Imports six layers deep that nobody documents…

What replaces them

A small, opinionated stack: a feature pipeline, a training script, a model registry, and a serving layer. Boring. Reliable. Reviewable.

CRM Reporting – Industrial Temps

Fri, 06 Sep 2024 00:00:00 +0000

Overview

A full refresh of CRM analytics for a recruitment agency: cleaner data, better KPIs, and faster monthly reporting.

Challenges

Inconsistent CRM fields → poor KPI accuracy (placements, fill rate, time-to-hire).
Slow manual exports → reporting lag and errors.
Limited visibility into drop-offs by stage (sourcing → placement).

What I Delivered

Data quality framework: mandatory fields, validation rules, audit queries.
ETL/ELT pipeline (Access + Power Query) → standardized reporting tables.
Power BI dashboards: placements, client performance, consultant scorecards.
Scheduled refresh from SharePoint & SQL; versioned datasets.

KPIs & Metrics

Placement rate, time-to-hire, fill rate, candidate source ROI.
Funnel drop-off by stage + reasons (disqualifications, no-shows).
Consultant activity vs outcomes (calls, interviews, starts).

Outcomes

3× faster month-end reporting (hours → minutes).
+22% data completeness (quarter-over-quarter).
Stakeholders self-serve common queries; fewer ad-hoc report requests.

Implementation Details

Source: CRM export + SQL mirror (where available)
Transform: Access/Excel (Power Query), data audits, normalization
Serve: Power BI reports with role-level security
Governance: Dataset owners, refresh alerts, change log

My Role

Requirement capture, data modeling, DAX measures, stakeholder training.
Built “single-click” monthly pack export for leadership.

Tooling

Dynamics 365 (or CRM export), Access, Excel/Power Pivot, Power Query, Power BI, SharePoint

Custom CRM Web Application

Fri, 06 Sep 2024 00:00:00 +0000

AI-Enhanced Social Healthcare Management System

Sun, 01 Sep 2024 00:00:00 +0000

AI-Powered AAC App

Sun, 01 Sep 2024 00:00:00 +0000

Overview

A modern AAC (Augmentative & Alternative Communication) app that helps non-verbal users express themselves faster.
It combines predictive text (TensorFlow.js), accessible UI, and a caregiver analytics dashboard for insight-led support.

Problem

AAC tools are often slow and cognitively heavy.
Caregivers lack visibility on what’s working (or not) during sessions.
Accessibility (WCAG) and privacy (GDPR) are frequently bolt-ons, not fundamentals.

Solution

Predictive AI suggests next words/phrases in context; improves with usage.
Accessible UI (high-contrast themes, large touch targets, offline first).
Caregiver dashboard: session logs, vocab usage, time-to-utter metrics.

Key Features

Next-word prediction via TF.js + custom tokenizer.
Personal vocab folders with pictograms; grid size & theme preferences.
Caregiver dashboard (Firebase + charts): top phrases, session durations.
WCAG 2.1 AA: keyboard/switch navigation, ARIA roles, focus states.
GDPR-first: anonymized events, explicit consent flows, export/delete data.

Architecture

Client: React / React Native (Kiosk mode on tablets)
Model: TensorFlow.js in-browser; optional server-assisted re-ranking
Storage: Firebase (Auth/Firestore/Storage)
Telemetry: Minimal, aggregated events; per-user opt-in

Outcomes

35–50% faster average time-to-utter for common phrases (pilot, n=7).
+40% caregiver satisfaction with session clarity (survey).
Reduced input errors with large targets & latency-friendly cache.

My Role

End-to-end build, model integration, WCAG compliance, privacy workflows.
UX research with caregivers; rapid iteration via weekly pilots.

Tech Stack

React, React Native, TensorFlow.js, Firebase (Auth/Firestore/Hosting), Tailwind, Charting

Screens / Media

Notes on Privacy & Safety

Anonymization by default, per-user data export, short retention windows.
Clear consent and session visibility for caregivers and admins.

Java-Based Ticket Machine System

Sat, 24 Aug 2024 00:00:00 +0000

Data Structures & Algorithms in Python

Sat, 15 Jul 2023 00:00:00 +0000

Cloud-Based Data Management System

Tue, 20 Jun 2023 00:00:00 +0000

Expense Tracker

Wed, 10 May 2023 00:00:00 +0000

Paul Martin McNeill's Portfolio

Building Production RAG Pipelines: Lessons from Real Deployments

Chunking is not preprocessing — it’s product design

Retrieval without grounding is just search

The eval harness is the real codebase

Why I Moved from Jupyter Notebooks to End-to-End ML Pipelines

The hidden cost of “it runs in my notebook”

What replaces them

CRM Reporting – Industrial Temps

Overview

Challenges

What I Delivered

KPIs & Metrics

Outcomes

Implementation Details

My Role

Tooling

Custom CRM Web Application

AI-Enhanced Social Healthcare Management System

AI-Powered AAC App

Overview

Problem

Solution

Key Features

Architecture

Outcomes

My Role

Tech Stack

Screens / Media

Web App Dashboard & Admin

Mobile App Screens

Firebase Backend & Data

Notes on Privacy & Safety

Java-Based Ticket Machine System

Data Structures & Algorithms in Python

Cloud-Based Data Management System

Expense Tracker