Ideas are easy.
Engineering is the work.

AI Systems

The Hard Part of AI Engineering: APIs, Retries and Idempotency

Engineering field notes on REST contracts, webhook delivery, job idempotency and ambiguous AI-provider outcomes in MailTrace and Instant Deck.

Architecture & product

Engineering note · Part 1

Building MailTrace: Turning Email Correspondence into a Searchable Evidence Workspace

How MailTrace connects correspondence, evidence, issues and saved answers without losing their source history.

Architecture & product

Engineering note · Part 2

Where Deterministic Code Ends and AI Begins in MailTrace

A practical boundary between model interpretation and the backend rules that protect source evidence.

Evidence & preparation

Engineering note · Part 3

Feature Engineering for Email Evidence: Threads, Participants, Dates and Requests

How useful email features combine semantic text with explicit context, and why missing context should remain visible.

Evidence & preparation

Engineering note · Part 4

Cleaning Email Text Without Changing the Original Evidence

Use derived text for retrieval while preserving original messages, source hashes and a route back to the evidence.

Embeddings & retrieval

Engineering note · Part 5

Using OpenAI Embeddings to Find Related Evidence in MailTrace

Where text embeddings fit in MailTrace, how source chunks are indexed, and what similarity does not prove.

Embeddings & retrieval

Engineering note · Part 6

Why Embedding Dimensions and Model Versions Matter in Production

How explicit embedding contracts prevent incompatible indexes and make migrations reviewable.

Embeddings & retrieval

Engineering note · Part 7

Hybrid Retrieval in MailTrace: Combining Semantic Search with Keyword Matching

Why evidence search needs both paraphrase matching and exact terms, with a small rank-fusion example.

Clustering & issue review

Engineering note · Part 8

Constrained Hierarchical Clustering for Email Evidence

How complete-link candidates and contextual constraints can reduce misleading issue merges.

Clustering & issue review

Engineering note · Part 9

Why Twelve Categories Do Not Mean Twelve Clusters

Keep broad navigation families separate from the number of concrete issues in a case.

Clustering & issue review

Engineering note · Part 10

Preventing False Merges: When Similar Emails Describe Different Requests

Why request references, dates and missing context must constrain semantic grouping.

Embeddings & retrieval

Engineering note · Part 11

Building Evidence-Grounded Q&A with RAG and the Responses API

How scoped retrieval, structured answer proposals and saved artifacts fit together in MailTrace Q&A.

Embeddings & retrieval

Engineering note · Part 12

A Citation Is More Than a Link: Validating AI Answers Against Source Evidence

Validate source scope, freshness and exact quotations before treating an answer as evidence-grounded.

Clustering & issue review

Engineering note · Part 13

Designing AI Suggestions That Users Can Review and Reverse

How proposal, apply and restore stages keep AI organisation changes understandable and recoverable.

Clustering & issue review

Engineering note · Part 14

The 69-Item Test That Exposed a Workflow Dead End

What a synthetic consolidation fixture revealed about bounded AI output, unassigned evidence and active issue visibility.

Production & efficiency

Engineering note · Part 15

Running AI Work in the Background with FastAPI and Persistent Jobs

Why a saved job lifecycle matters more than a spinner when AI work is slow, retryable and paid.

Production & efficiency

Engineering note · Part 16

Reducing Repeated AI Work with Versioned Indexes and Saved Results

How reuse, explicit invalidation and saved artifacts reduce latency without hiding stale evidence.

ML experiments & evaluation

Engineering note · Part 17

Testing an AI Application: What Mocked Tests Prove—and What They Leave Unanswered

Separate deterministic regression tests, provider integration checks and real evidence quality evaluation.

Architecture & product

Engineering note · Part 18

Making AI Processing Understandable in the MailTrace Interface

Turn backend states into clear actions, saved progress and honest explanations for people working with evidence.

ML experiments & evaluation

Experiment design · Part 19

Can Logistic Regression or Naive Bayes Classify MailTrace Messages Cheaply?

A proposed baseline experiment for narrow email triage, with a scikit-learn pipeline example.

ML experiments & evaluation

Experiment design · Part 20

Random Forest, XGBoost, LightGBM and CatBoost for Email Triage

Where tree ensembles might help structured email triage, and how to compare them without inventing a winner.

Clustering & issue review

Experiment design · Part 21

K-Means, DBSCAN and Hierarchical Clustering: Which Fits Evidence Organisation?

Compare clustering assumptions against the actual cost of merging or separating case evidence.

Clustering & issue review

Experiment design · Part 22

Could Gaussian Mixture Models Represent Ambiguous Evidence Groups?

Explore soft cluster membership without mistaking model responsibilities for evidence certainty.

Embeddings & retrieval

Experiment design · Part 23

Reranking Retrieved Evidence Before Asking the LLM

A proposed second-stage relevance check for MailTrace retrieval, with clear limits on cost and authority.

ML experiments & evaluation

Experiment design · Part 24

Cross-Validation for Email Data: Preventing Thread and Case Leakage

How to evaluate email models without training on near-duplicates of the evidence used to test them.

ML experiments & evaluation

Experiment design · Part 25

Learning from Reviewer Corrections in MailTrace

Turn reviewed decisions into a governed evaluation dataset before using them to train another model.

Evidence & preparation

Experiment design · Part 26

Can a Small Classifier Separate Signature Assets from Useful Evidence?

Combine attachment signals without excluding a useful file solely because it is small.

ML experiments & evaluation

Experiment design · Part 27

PCA and t-SNE for Exploring MailTrace’s Embedding Space

Use embedding visualisations to investigate patterns without treating a two-dimensional picture as a clustering verdict.

ML experiments & evaluation

Experiment design · Part 28

Isolation Forest and One-Class SVM for Evidence-Import Anomalies

Explore anomaly detection for import quality while keeping unusual evidence distinct from invalid evidence.

ML experiments & evaluation

Experiment design · Part 29

When Would Fine-Tuning Improve MailTrace?

Decide whether a measured failure belongs to retrieval, prompting, validation or model training before fine-tuning.

Production & efficiency

Experiment design · Part 30

Choosing Smaller Models for Routine Tasks in an Evidence Application

A task-specific routing experiment that weighs quality, latency and cost without using model size as a safety rule.

Architecture & product

Architecture proposal · Part 31

Using Pretrained AI Versus Training Models with TensorFlow or PyTorch

Explain where deep-learning frameworks could fit without implying that every AI product needs its own neural-network training stack.

Architecture & product

Architecture proposal · Part 32

Do MailTrace’s Workflows Need LangGraph, CrewAI, AutoGPT or Bee?

Evaluate orchestration frameworks against durable jobs, explicit tool permissions and the existing application workflow.

Evidence & preparation

Engineering note · Part 33

From PDF to Searchable Evidence: Text Extraction, OCR and Multimodal Processing

Distinguish a stored file from searchable text, and expose unsupported media processing honestly.

Architecture & product

Architecture proposal · Part 34

Designing One Evidence Backend for Web and Android

Share evidence rules and saved jobs across clients while giving Android an interface suited to mobile use.

Architecture & product

Architecture proposal · Part 35

What a Year of MailTrace Iterations Taught Me About AI Product Architecture

Why consolidation needs maintained capability owners, explicit contracts and workflow tests rather than another collection of copied modules.

RAG & AI Evaluation

How to Fine-Tune and Optimise RAG Systems for High-Stakes Decision Support

A practical guide to improving RAG systems through embedding fine-tuning, FAISS optimisation, hybrid retrieval, generative-model tuning, evaluation and continuous monitoring.

RAG Fundamentals

RAG Decision Support Explained: Retrieval, FAISS and Grounded Generation

A clear explanation of how decision-support logic fits into RAG, what FAISS does, and why RAG can provide better domain-specific context than a standalone LLM.

GraphRAG & Knowledge Systems

RAG + Knowledge Graphs: How Graph-Aware Retrieval Improves Context Relevance

How combining Retrieval-Augmented Generation with knowledge graphs improves entity linking, relationship-aware retrieval, multi-hop context and grounded answers.

Agentic RAG

Agentic RAG: Multi-Step Reasoning with Retrieval and Generation

How agentic RAG extends standard retrieval-augmented generation with planning, iterative retrieval, tool use, validation and multi-step reasoning.

Agentic AI

How to Design Reliable Agentic AI Workflows

Practical lessons from building Open Tools: bounded agents, explicit permissions, durable task memory, validation gates and deployment checks.

RAG & Search

RAG for Evidence-Heavy Businesses: Source-Linked Retrieval That People Can Trust

How to design RAG and semantic retrieval for evidence, compliance and document-heavy workflows using source boundaries, pgvector and auditability.

AI Product Design

Human-in-the-Loop AI for Sensitive Workflows

Why AI outputs should often remain drafts until reviewed, and how structured schemas, confidence and confirmation improve trust.

AI Product Engineering

From AI Prototype to Production Product: What Changes After the Demo Works

A practical production checklist from DDDecks.ai: observability, error contracts, security, failure handling, data boundaries and deployment.

CRM & Product Design

How to Design a Relationship-Intelligence CRM Instead of Another Contact Database

Product lessons from Founder CRM: goals, people, companies, opportunities, warm paths, scoring and next actions.

Generative AI

AI Document Generation Beyond Prompt Engineering

Lessons from Instant Deck on evidence extraction, research, memo synthesis, narrative planning, visual planning and controlled rendering.

Vector Search

pgvector Semantic Search: Production Lessons from Evidence Retrieval

Practical guidance on using PostgreSQL and pgvector for semantic search, with metadata filters, ownership boundaries and retrieval evaluation.

AI Evaluation

Structured Outputs and LLM Evaluation for Production AI Products

How schemas, validation, critique, repair and regression testing improve reliability in LLM-powered workflows.

Technical SEO

SEO-Safe WordPress to SvelteKit Migration: What I Would Protect First

Lessons from migrating a large content site: route inventories, canonical URLs, indexing gates, media migration and deterministic content.

AI Strategy

Building AI Products Around Real Business Workflows

A cross-project framework for deciding where AI belongs: workflow mapping, structured context, model boundaries, human review and measurable outcomes.

Parallel AI

Parallelising Embedding Generation for Large Document Collections

How to design embedding pipelines that scale with batching, partitioning, retries, idempotency and cost-aware concurrency.

HPC & ML Systems

What Slurm and HPC Scheduling Teach Us About Production ML

Lessons from scientific HPC that apply directly to batch inference, evaluation, document processing and AI pipelines.

Machine Learning

DBSCAN vs K-Means: Choosing a Clustering Model for Real Product Questions

A practical comparison of DBSCAN and K-means based on safeguarding and marketplace segmentation projects.

Responsible ML

Synthetic Data for Sensitive Machine Learning: What It Can and Cannot Prove

How synthetic data can support privacy-preserving prototyping without pretending simulated labels are real-world evidence.

ML Evaluation

Machine Learning Evaluation Beyond Accuracy

Why precision, recall, F1, calibration, failure costs and operational thresholds matter more than one headline metric.

Data Science

Feature Engineering for Longitudinal Behavioural Data

Practical ways to turn event-level histories into trends, volatility, rolling windows and cohort-comparable features.

Data Products

Designing an R Shiny Dashboard Around Machine-Learning Decisions

How to integrate models into interactive dashboards without overwhelming non-technical users.

Human-Centred AI

Designing Risk Dashboards for Human Review

Human-centred principles for alerting, triage and machine-learning outputs in sensitive workflows.

RAG Evaluation

How to Evaluate a RAG System: Retrieval, Grounding and Product Quality

A practical RAG evaluation framework covering retrieval recall, citation correctness, unsupported claims, latency and cost.

Search Engineering

Hybrid Search: When Vector Similarity Is Not Enough

Why production search often combines semantic retrieval, lexical matching and metadata filters.

RAG & Search

Reranking for RAG: Improving Relevance After Retrieval

How reranking can improve retrieval quality without replacing your vector database.

Agentic AI

Observability for AI Agents: What to Log Without Creating a Privacy Problem

A practical observability model for agent runs, tool calls, retries, failures and sensitive data boundaries.

AI Evaluation

Regression Testing Prompts and LLM Workflows

How to stop prompt edits and model upgrades from silently breaking production behaviour.

Production AI

Model Routing: Balancing AI Quality, Cost and Latency

How to route requests across models based on task complexity, risk and response-time requirements.

Production Engineering

Structured Logging for Production AI Applications

How request IDs, event names and redaction make AI systems easier to debug and safer to operate.

AI Systems

Background Jobs for Long-Running AI Workloads

Why document processing, report generation and multi-step AI tasks should often move out of the request-response cycle.

AI Product Architecture

Designing PostgreSQL Data Models for AI Products

Why durable AI products need explicit entities, source records, artifacts and model-run metadata—not just chat history.

Technical Product

Entity Modelling for AI Product Managers

How entities and relationships turn vague AI features into buildable product specifications.

Technical Product Management

Roadmapping AI Products Under Uncertainty

How to prioritise platform capability, configuration and experiments when model capability and customer needs move quickly.

Enterprise AI

AI POC to Enterprise: A Practical Product Checklist

What has to change when an AI proof of concept becomes a real multi-user product.

Data Visualization

Tableau Dashboard Design Lessons from 97,224 Airbnb Listings

What a large London marketplace dataset taught me about aggregation, filtering, coordinated views and analytical storytelling.