How I identified manual work, built AI-operated systems around it, and created a repeatable way to automate more of the organization’s operations.
Initiative ownership
I started the initiative, then built the systems.
Since March 2026, I have been finding manual processes across research and content operations and turning them into workflows that can be run through AI with minimal input.
I initiated the work and chose the projects. For each one, I learned how information moved between people and platforms, gained access to the APIs, and decided how far the process could be automated safely.
The method I developed
From manual process to AI-operated system
01
Find the work
I looked for recurring processes where analysts were moving the same information between systems, repeating checks, or rebuilding the same output by hand.
02
Learn the process
I mapped the inputs, decisions, exceptions, and failure points. I also worked through API access and the constraints of each platform.
03
Design the boundary
I decided what the AI could handle, what belonged in deterministic code, and where a person still needed to review or approve the work.
04
Build and test it
I built the integrations, AI instructions, validation harnesses, dry runs, and recovery paths needed to use the workflow on real work.
05
Make it reusable
I packaged the work with configuration, skills, runbooks, decision records, and version control so the system could be maintained and handed off.
Project evidence
Each project, one by one
01
AI evaluation
STAR Awards judging system
How the workflow connects
Inputs
APIClickUp API
Application attachments and award category
FilesApplication documents
Text and rendered page images
Processing
CodePython
Ingest and check application content
AIOpenAI API
Judge panel produces scores and feedback
Outputs
DocumentsReviewer briefs
Evidence and scores for review
DocumentsPersonalized feedback
Applicant reports and winner case studies
People review the evidence and make the final award decisions. Human scores and referee models are used in the separate validation study.
Problem
Award judging depended on hundreds of hours of manual review, with no tested AI evaluation system in place.
What I built
I designed the judging architecture, criteria, scoring personas, processing pipeline, and validation study. When the first referee design rewarded polished writing instead of evidence, I caught the failure and rebuilt the test.
Company value
The live process moved from 330 hours of human grading to 10 hours of human review while preserving human oversight for the final decisions.
Claude CodeLLM evaluationPythonStatistical validationHuman review
Validate the specification and execute API changes
Outputs
APIGoogle Sheets
Tracking workbook
APIAlchemer
Built survey
APIContentful
Poll listing draft
APIClickUp
Project task and tracking links
The operator approves checkpoints before the pipeline proceeds. New Contentful entries remain drafts for review.
Problem
Producing a poll required a long checklist of updates across ClickUp, Google Sheets, Alchemer, and Contentful.
What I built
I mapped the full lifecycle, connected each API, and built a Python pipeline with an AI operator layer. The user supplies a ClickUp task ID, answers a few questions, and approves the checkpoints that still require judgment.
Company value
A first-time operator completed setup and build with no manual system edits. The repetitive execution now happens inside the workflow.
Claude CodePythonClickUpGoogle SheetsAlchemerContentful
03
Data quality
Benchmark validation engine
How the workflow connects
Inputs
FilesBenchmark CSV exports
Tall and wide response tables
ConfigPractice configuration
Validation rules and instrument settings
Processing
CodePython · pandas
Check outliers, zeros, and missing values
Outputs
FilesExcel workbooks
Flagged values and review decisions by practice
Analysts decide whether to keep, remove, or correct each flagged value. This validation step leaves source files unchanged.
Problem
Each research practice had its own version of the validation logic, which made fixes and updates difficult to keep consistent.
What I built
I replaced the separate scripts with one configuration-driven Python engine. I kept member data out of the AI coding workflow and tested the system with synthetic fixtures and reconciliations against manual counts.
Company value
The research practices now share one validation system, and a new practice can be added through configuration instead of another copy of the code.
Claude CodePythonData validationSynthetic fixturesConfiguration
04
Research tools
Read-only benchmark instrument browser
How the workflow connects
Inputs
Read-only APIContentful Delivery API
Published instruments, modules, and questions
InputAnalyst request
Practice, question type, or module selection
Processing
CodePython
Traverse instruments and filter questions
Outputs
FileExcel
Selected instrument content, organized by module
The client is read-only: browsing and export cannot change the published content.
Problem
Published benchmark questions and modules lived in Contentful, where analysts could not easily browse, filter, or export them.
What I built
I built a tool that reads the instrument structure through the Contentful delivery API, filters the questions, and exports selected modules to Excel.
Company value
Analysts can retrieve the material they need without navigating the CMS or giving the tool permission to change the source content.
Claude CodePythonContentful APIExcelRead-only access
05
Content operations
Benchmark instrument updater
How the workflow connects
Inputs
FileExcel
Question, practice, and metric definitions
Processing
CodePython
Read rows and create update payloads
Outputs
Write APIContentful API
Created or updated instrument definitions
FileRow-level log
Success or failure for each input row
This is a write-enabled tool. The result log records which rows succeeded or failed.
Problem
Loading question, practice, and metric definitions into the CMS required repeated manual entry.
What I built
This was one of my first write-enabled API projects. I built a Python tool that reads the definitions from Excel, creates or updates them in Contentful, and records the result for every row.
Company value
The instrument setup can be run as a structured load with a record of what succeeded or failed.
Claude CodePythonExcelContentful APIResult logging
06
Content automation
Survey report publishing workflow
How the workflow connects
Inputs
FileAlchemer export
Completed survey results
Processing
WorkflowAlteryx
Filter and transform results
AIOpenAI API
Draft chart captions and an executive summary
Outputs
DraftsContentful API
Research-report entries and generated text
The analyst reviews the drafts and controls publication.
Problem
Publishing survey results required transforming the data, writing chart captions and a summary, and entering the material into the CMS.
What I built
I built an Alteryx workflow that prepares the results, calls the OpenAI API for draft captions and an executive summary, and creates the Contentful entries as drafts.
Company value
The workflow performs the repeated preparation and content entry. The analyst reviews the generated drafts before publication.
Claude CodeAlteryxAlchemerOpenAI APIContentful API
07
API integration
Conference content sync and author mapping
How the workflow connects
Inputs
APICvent API
Conference sessions and speaker information
ReferenceContentful records
Existing authors and archived session matches
Processing
CodePython sync
Filter sessions, match authors, and prepare updates
Outputs
DraftsContentful API
Conference-session drafts
Author matches are checked before production writes. The sync preserves fields maintained by the content team.
Problem
Conference sessions created in Cvent had to be entered again in Contentful, and every speaker needed to be matched to the correct author record.
What I built
I built a one-way Python sync, added the matching logic, and checked the author results against archived sessions before any production write.
Company value
The workflow creates reviewable CMS drafts while preserving fields maintained by the content team and surfacing mismatches for investigation.
Claude CodePythonCvent APIContentful APIValidation
08
Intake automation
Speaker intake pipeline
How the workflow connects
Inputs
Survey dataAlchemer
Speaker application responses
Processing
CodePython
Clean responses and map answers to task fields
Outputs
APIClickUp API
Speaker application tasks and custom fields
The conference team manages the resulting applications in ClickUp.
Problem
Speaker applications collected in Alchemer had to be cleaned and re-entered in the conference team’s ClickUp tracker.
What I built
I replaced the prior intake workflow with Python scripts that clean the survey responses and map each answer to the correct ClickUp field.
Company value
Applications move from the survey platform into the team’s working system without being re-keyed by hand.
Claude CodePythonAlchemerClickUp APIData cleaning
09
Data engineering
Financial index data pipeline
How the workflow connects
Inputs
APIFinancial Modeling Prep API
Financial statement data
Processing
WorkflowAlteryx
Transform data for the recurring index refresh
Outputs
WarehouseBigQuery
Financial index data tables
ReportingTableau
Dashboards and outlier review from the warehouse
The pipeline owner reviews data quality and outliers during the quarterly refresh.
Problem
The recurring financial index refresh depended on manual vendor-workbook entry and a separate set of quality checks.
What I built
I migrated the process to an API-driven Alteryx pipeline feeding BigQuery, with outlier review in Tableau. I also owned the refresh process, documentation, and checks.
Company value
The quarterly refresh runs through a maintained pipeline with the data checks attached to the same operating process.
Claude CodeAlteryxBigQueryTableauAPIs
10
AI enablement
AI workflow handoff system
How the workflow connects
Inputs
InputProcess knowledge
Rules, credentials boundaries, and decisions
FilesProject files
Configuration, tests, and runbooks
Processing
AI instructionsClaude Code skills
Encode how to operate each workflow
Outputs
HandoffVersioned project repository
Reusable skills, YAML, and documentation in GitHub
HandoffOperator workflow
Instructions for running, reviewing, and recovering
This is the operating framework used across projects. Each workflow documents where human approval is required.
Problem
An automation is difficult to sustain when only its builder knows how to run it, test it, or recover from an error.
What I built
I created reusable skills, configuration files, project instructions, runbooks, decision records, credential boundaries, and GitHub repositories. Together, they teach the AI how to operate each workflow and tell the user where approval is required.
Company value
The system gives other operators a documented route into the work and gives future maintainers a record of why the workflow was built this way.
Claude CodeReusable skillsYAMLGitHubRunbooksDecision records
11
AI-assisted frontend work
Alchemer survey styling automation
How the workflow connects
Inputs
Web pageAlchemer survey
Page markup and stylesheet loading order
InputBrand template
Required appearance and styles
Processing
InspectionChrome DevTools
Inspect where styles are overridden
AI instructionsClaude Code skill
Apply the tested platform-specific CSS rules
Outputs
CSSAlchemer
Branded survey styling
HandoffReusable skill
Platform rules for future surveys
This is a CSS and platform-behavior workflow, rather than an API data sync.
Problem
Alchemer’s page structure and stylesheet loading order caused custom survey styles to be overridden.
What I built
I inspected the live survey DOM and loading order, found where the CSS needed to run, and packaged the platform-specific rules and brand template into a reusable AI skill.
Company value
Future survey styling starts with the tested platform rules instead of rediscovering the same undocumented behavior.
Claude CodeCSSChrome DevToolsAlchemerReusable skills
12
API tooling
ClickUp custom-field tooling
How the workflow connects
Inputs
APIClickUp list schema
Source fields and target list configuration
Processing
CodePython
Create, copy, or update custom-field schemas
Outputs
Write APIClickUp API
Consistent custom fields across lists
The field structures support intake workflows, including the related Alchemer response sync.
Problem
Intake workflows depended on matching custom-field structures across several ClickUp lists.
What I built
I built Python tools that create, copy, and update the field schemas and support the related survey-response sync.
Company value
The field setup can be repeated through code, which reduces manual configuration and keeps the intake structures consistent.
Claude CodePythonClickUp APIAlchemer
13
Conversational automation
Meeting notes to ClickUp tasks
How the workflow connects
Inputs
TextMeeting notes
Follow-up actions and task details
Processing
AIClaude Code skill
Identify tasks and ask for missing information
Outputs
TasksClickUp
Structured follow-up tasks
The conversation supplies missing details before tasks are created.
Problem
Meeting notes contained follow-up work that still needed to be translated into structured project tasks.
What I built
I created conversational automations that read the notes, identify the task details, ask for missing information, and create the ClickUp tasks.
Company value
The workflow provides a consistent route from a conversation record into the project system.
Claude CodeReusable skillsClickUp
14
Survey reporting
Poll reports and survey dashboards
How the workflow connects
Inputs
InputCompleted poll results
Data from the closed poll
Processing
AI instructionsClaude Code skills
Guide report and dashboard preparation
Outputs
OutputPoll reports
Prepared results for reporting
OutputSurvey dashboards
Prepared dashboard material
These reusable skills support the reporting stage of the survey lifecycle.
Problem
The survey lifecycle needed a repeatable process for turning completed poll results into reports and dashboards.
What I built
I created AI automations that guide the preparation of poll reports and survey dashboards as part of the larger survey workflow.
Company value
The reporting steps are documented in an executable workflow that can be reused after each poll closes.
Claude CodeReusable skillsSurvey reportingDashboard preparation
What this added
A reusable AI automation capability
I built a way to keep finding, evaluating, and automating suitable work. The company gained working systems, a safer operating pattern for AI, and the documentation and infrastructure needed to extend the approach.