AI that supports a workflow
Large language model development and integration for document review, analysis, triage, and decision support. Clear boundaries matter: an AI suggestion stays a suggestion.
Independent contractor · Construction + utilities
Practical AI for the work behind the work.
I build clear, useful AI and data workflows for people who work with real equipment, real records, and real deadlines. This first-pass cathodic protection review shows the approach: make the information easier to inspect, keep the original evidence visible, and put decisions in the hands of the specialists who own them.
First-pass projectReview framework awaiting specialist approval
From messy input to a useful decision
01Understand the sourceKeep the company's original records and labels intact.
02Show the logicDocument each score, coverage rule, and uncertainty flag.
03Make review easierGive the specialist a focused queue and a clear reason to look.
Human judgment stays in the loop.
What I do
Build tools that fit the job.
I work independently, from the first messy workbook through a usable handoff. The goal is a process people can actually understand and use.
From scattered records to clarity
Spreadsheets, repeatable scoring logic, review queues, quality checks, and readable deliverables built around how field and office teams actually work.
Computers built for the work
I also build and configure computers. The software and hardware should work together instead of getting in your way.
A new tool in the stack
Fast, structured decisions with uncertainty in view.
TypeSafe Jev is a new early-access model built for typed decisions with probabilities and confidence. In this project, it gave each de-identified record an advisory review-urgency score while keeping its answer distribution available for inspection.
That fits repeatable triage well. A frontier model such as Codex can organize the workflow and explain the results; deterministic code owns the rubric; Jev handles a narrowly defined decision across many records. A fine-tuned model running locally could join a future version of this stack after its own evaluation on approved data.
TypeSafe's own published workflow evaluations report strong speed and accuracy for this kind of structured decision. We have not measured a CP-specific speedup or validated Jev's urgency against specialist-labeled station outcomes.
Featured first pass
A better way to review CP station records.
A transparent draft rubric and review queue for cathodic protection data, built from a supplied workbook and a CP3's explanation of the measurements.
155location records reviewed
81provisional grades supported by the draft rules
74left unscored where core evidence was insufficient
111rows surfaced with at least one review flag
The point: the draft separates a calculated condition grade from the company's original Pass/Fail result, and shows where a specialist needs to resolve missing or conflicting evidence. These are aggregate project figures; station-level records are shared privately.
The proposed CP rubric
What the grade actually means.
A CP3's column explanation supplies the measurement meanings and the −850 mV and 100 mV reference points. The factor weights and A–F bands are our proposed review framework, ready for specialist calibration and approval.
Seven weighted factors
| Factor | Weight |
|---|---|
| Instant-off potential | 35% |
| Polarization decay | 25% |
| Isolation status | 15% |
| Current adequacy | 10% |
| On-potential consistency | 5% |
| IR contribution | 5% |
| Resistance peer consistency | 5% |
Condition grade bands
| Grade | Score |
|---|---|
| A | 80–100 |
| B | 60–<80 |
| C | 40–<60 |
| D | 20–<40 |
| F | 0–<20 |
At a boundary, the score enters the higher grade: 20 is D, 40 is C, 60 is B, and 80 is A. Missing core evidence is Unscored, never an automatic F.
How this guides review: eligible rows can be sorted from lowest condition score upward; the factor detail shows which weighted inputs pulled a score down. Unscored or conflicting records stay in a separate review queue. The company's original Pass/Fail remains visible, while Jev's urgency is a separate advisory signal for when a specialist should look.
About
Hi, I'm Zach.
I'm a solo independent contractor working at the intersection of AI, large language models, construction, and utilities.
I like practical problems: a workbook nobody trusts, a process that takes too long, or a pile of information that needs a clear path to review. I can build the data workflow, explain the logic, and hand over something the next person can use. I build computers too, because good tools need a solid place to run.
How I work
Plain language. Visible assumptions. Useful outputs.
- Start with the actual source material and subject-matter expertise.
- Keep original outcomes separate from proposed analysis.
- Make uncertainty visible instead of hiding it behind a score.
- Deliver files and explanations people can inspect and revise.
What comes next
A solid first pass, ready for the next conversation.
The CP framework is a draft for engineering review. The rubric, source questions, and toolchain are documented here so a specialist can challenge the assumptions and shape the final version.