About
One engineer, four disciplines, no hand-offs.
Chris Bacon builds AI systems, data infrastructure, and custom software — and, unusually for an engineer, does the technical SEO and analytics work that makes any of it get found.
DataSync Solutions is a one-person-deep engineering studio, expanded with trusted collaborators when a project needs it. You get senior judgement on every line rather than a junior behind an account manager.
The range is not a claim, it is a track record you can read. A matching engine that indexes 400,603 job postings and explains every score it produces. A modular LoRA fine-tuning pipeline, and the adversarial audit that caught it memorising instead of reasoning. A cross-platform product where two native applications share one deterministic core so their behaviour cannot drift.
The unusual part is the last mile. Most engineers hand over a repository and leave; most agencies market something they could not have built. Doing both means the model knows about your data, the application is designed around the model, and the analytics are wired in from the first line rather than bolted on at launch.
Client engagements are confidential. Where client delivery appears in the portfolio it is anonymized — what was built and the scale where disclosure is safe, never a client name, their data, or the method. Everything else shown on this site is our own work, which is why the numbers and the failures can both be shown.
The background
Ask the shell.
The full history is here, and so is the PDF — but neither is sitting in a folder waiting to be harvested. Type a command.
The PDF is not a file sitting in a public folder and there is no link to it on this page. Typing resume answers a short-lived challenge and the server streams it back — which a person does in one keystroke and a crawler does not do at all.
How we work
Four rules, all of them inconvenient.
Show the work
Every ranking, score, and recommendation on this site explains why it produced the answer it did. A system that cannot be audited cannot be trusted with a decision that matters.
Name the limit
A feature in one of our projects was renamed mid-build because the data could not support the original claim. Saying what a system cannot do is cheaper than defending an overclaim to a client later.
Break your own thing first
Our fine-tuning pipeline was audited adversarially and 84 issues came back. A system nobody has tried to break is not evidence that it works.
The simplest thing that works
A hosted API before a fine-tune. Deterministic scoring before a model. Plain automation before AI. Half of good AI consulting is talking someone out of a model they do not need.
Active research
Work in progress, labelled as such.
Two design efforts are currently specification only. They are listed because the thinking is real, and flagged because the code is not — the same rule applied everywhere else on this site.
Model capacity inference
Profiling a black-box language model into a capability manifold using structured probe batteries and behavioural signatures — no weights, no GPU, no side channels, only the semantic content of responses.
Outbound network policy agent
An eBPF agent that observes outbound network behaviour per workload, generates confidence-tiered least-privilege policies, and gives administrators a simulation path before anything is enforced.
Related
semantic rankingContinue with another article in this topic.
- Related resourcecosine match 0.66Selected work: built, tested, measuredClosest content match: Selected work: built, tested, measured.Open resource
- Interactive democosine match 0.64The Lab — live demosClosest content match: The Lab — live demos.Open demo
- Related articlecosine match 0.61The development process I run, and where it has failedClosest content match: The development process I run, and where it has failed.Read article
Want to work together?
A free 30-minute call. We'll tell you honestly whether we're the right fit and what the work would take.