Services · AI workflows
AI workflows
AI can code thousands of responses in minutes, but it’s not easy to know if you can trust the results. I build workflows that are checked against human coding before they run on your full dataset so you can have faith in your findings.
Workflow development and training
Can our team run this ourselves?
I build an LLM-assisted workflow for your data and questions, validate it against human coding, and train your team to run it on new data
- You get
- A documented workflow, agreement statistics, and training for your team
- Good fit
- Research, evaluation, and economic development teams that collect a lot of open-ended text
Large-scale text analysis
What are all those qualitative data signaling?
For survey responses, transcripts, documents, or scraped data, I build an LLM-assisted workflow for your dataset. The model’s coding must agree with mine on a random sample (Cohen’s κ > 0.7) before it touches the full dataset. Data handling is agreed in writing before any data change hands.
- You get
- A codebook, a coded dataset, themes with representative quotes, agreement statistics you can cite, and a documented workflow you can rerun
- Good fit
- Survey comments, interview archives, public comment records, and job postings
How it works
Scope
What questions the findings should answer, and what a correct answer looks like
Build
A workflow tuned to your data, with edge cases worked out first
Validate
The model’s coding must agree with human coding on a random sample (Cohen’s κ > 0.7) before it runs on everything
Hand off
Documentation, agreement statistics, and training so your team can rerun it on new data
What you’ll have at the end
The form depends on who will run it:
A ready-to-run notebook
For teams with an analyst who works in Python
A tested prompt set
For the AI tools you already use, with step-by-step instructions
A simple upload-and-run tool
For teams with no technical staff who want something they can use right away
A documented process
For analysts who want to follow it with the tools they have
Whatever the form, it is validated against human coding, documented, and built so your team can keep using it
Example
An automated program evaluation report
One workflow turns a survey export into a two-page report: ratings, coded themes with representative quotes, and recommendations tied to the evidence. The same workflow reruns each cohort.
Illustrative data, not a real program
See the sample report →How engagements work
Setup and training
I build and validate the workflow, document it, and train your team to run it
Ongoing support
Periodic check-ins and adjustments as your data or questions change
Full analysis
I run the workflow on your data and report the findings