On Delinquent Data

It is of the highest importance in the art of detection to be able to recognise, out of a number of facts, which are incidental and which vital. Otherwise your energy and attention must be dissipated instead of being concentrated.

— Arthur Conan Doyle (The Memoirs of Sherlock Holmes)

"It began with a rather unremarkable observation: data is rarely as reliable as we would like to believe." That's how an LLM would commence, but let me give you the action at the outset: I recently built Sentinel.

Sentinel is a Python-based data reliability and incident investigation platform that monitors datasets for quality issues, schema changes, distribution drift, and anomalies. Combining statistical analysis and configurable validation rules with persistent incident management, background processing, and a REST API, Sentinel also features an optional LLM-powered investigation layer designed to generate evidence-grounded explanations without replacing deterministic analysis. The underlying stack comprises Polars, PostgreSQL, FastAPI, Celery, and scikit-learn.

Impressive, if I do say so myself. I should also say that one person can only test so much, so I would appreciate others using it and sharing feedback. Marketing aside, though, there is an origin story to this project. In this economy more turbulent than the turbulent term of Tyke Tiler, I understandably spend some time on LinkedIn, where I notice a lot of job titles in my "field" clustering around overlapping yet distinct sets of skills. Job titles themselves are often awfully ambiguous too, but I digress. Perhaps other fields observe a similar pattern. Data analyst, data scientist, data engineer, AI/ML engineer, and so on. Once in a while there's an AI-generated post on how shades of gray are here at play. But not today, I may say, as here I inveigh against a trend so passé, and I won't delay till this work sees the light of day. Uh, maybe all that was merely a ricochet and didn't stay. Anyway, I just wanted to highlight how Sentinel combines components from all those similar-sounding roles.

Then did I only want to be eligible for all those roles with a single stroke? No, good bloke. Not only. I believe many companies would benefit from a product like Sentinel. Here's a common scenario I, and many, have personally encountered. A data pipeline runs successfully while quietly producing bad data. A disappearing column, a shifting distribution, duplicate identifiers. Worse, downstream systems continue consuming the results none the wiser. Now, imagine a fairy tale. A system ingests datasets, preserves their versions, profiles their contents, detects anomalies and changes, and turns its findings into incidents that could be investigated and tracked. That is Sentinel. Of course, the project grew in layers. First came ingestion and profiling. Then configurable quality rules and statistical detection. Eventually, there were historical baselines, drift analysis, incident policies, persistent evidence, background workers, an API, and an operator dashboard. The learning curve was just shy of a Victorian roof.

At any rate, we live in big 2026, so let's talk about AI. Explore Sentinel's code, and you can see it benefited from AI assistance. More importantly, AI is part of the product itself. That said, the idea is not to throw raw data into the black box that is AI and—presto—you have a theory of everything (read: data analysis). Admittedly, that could work, and possibly work better, but a bit of explainability does not hurt either, and there is always the ever-so-slim chance of hallucination. Therefore, deterministic analysis is a major chunk of this work. A plausible counter to this approach is that humans learn through both reason and experience, a priori and a posteriori. Sure, we do not yet have a formal account of AI's internal reasoning, but experience offers increasingly convincing reasons to believe that AI knows what it's doing. And I do not reject that counter.