03 / AI-ASSISTED DECISION SYSTEM
Building AI around judgment, not instead of it
My job search was creating a familiar kind of operational problem: hundreds of recommendations arriving from fragmented sources, inconsistent information, duplicate listings, subjective evaluation, and repetitive application work.
I built an AI-assisted operating system to consolidate those inputs, enrich and normalize the data, evaluate opportunities against an explicit decision framework, and generate tailored application materials—while keeping consequential career decisions and final approval with me.
MY ROLE / System design · AI workflow architecture · decision framework · data normalization · human-in-the-loop
~300 / week
Job recommendations processed
Unique opportunities evaluated (and counting)
~100 → ~20 min
Review time per 20 opportunities
~3 - 5 min
Human review time for tailored application materials
The goal was never to automate the decision to apply. It was to automate the processing required to make that decision well.
THE PROBLEM
The hard part wasn’t finding jobs. It was processing enough information to make good decisions.
The recommendations were already there. The problem was what happened after they arrived.
Listings came from multiple sources with different fields, missing information, duplicate postings, inconsistent salary and location data, and varying levels of detail. Before I could decide whether a role was worth pursuing, I had to clean and interpret the inputs first.
That meant a lot of time was being spent on information processing before I ever reached the part that actually required judgment.
PROCESSING WORK
· Consolidate sources
· Fill missing data
· Normalize fields
· Remove duplicates
· Compare requirements
DECISION WORK
· Consolidate sources
· Fill missing data
· Normalize fields
· Remove duplicates
· Compare requirements
AUTOMATE MORE OF THIS → PRESERVE MORE OF THIS
~100 min → ~20 min per 20 opportunities
Reducing the repetitive processing created more room for actual evaluation.
The goal wasn’t to make the decision automatic. It was to make the decision easier to reach.
Once I separated processing from judgment, I could start designing the system around that boundary.
V1 / INGEST
Give every opportunity the same shape.
Job recommendations arrived from multiple sources, each with different fields, formats, and levels of detail.
Before I could compare opportunities consistently, I needed to turn those fragmented inputs into a common record. I created a structured ingestion layer so every role entered the system with the same core fields and could move through the same evaluation workflow.
Indeed
BuiltIn
STRUCTURED OPPORTUNITY RECORD
Company
Role
Location
Compensation
Requirements
Source
WHY THIS CAME FIRST
AI could not evaluate opportunities consistently if the same concept appeared differently—or was missing entirely—across sources.
A shared structure gave every later layer of the system something stable to work from.
The first version didn’t make better decisions. It made consistent decisions possible.
Once every opportunity had a common structure, the next problem became obvious: the records were consistent, but the underlying information was not.
V2 / ENRICH + CLEAN
Consistent records still needed better information.
Once every opportunity had the same structure, the next problem was completeness and consistency.
Listings often arrived with missing compensation, unclear work arrangements, inconsistent location formats, stale availability, or duplicate postings from multiple sources. I added an enrichment and cleanup layer to improve those records before they reached evaluation.
RAW RECORD
Salary: —
Location: Los Angeles / Remote?
Work arrangement: —
Status: unknown
Source: LinkedIn
Possible duplicate
ENRICH + CLEAN
FILL · NORMALIZE · VERIFY · DEDUPE
STRUCTURED RECORD
Salary: $X - $Y
Location: Los Angeles, CA
Work arrangement: Remote
Status: Active
Source: LinkedIn
Duplicate: No
01 / ENRICH
Fill the gaps that matter
Add missing details such as compensation, work arrangement, availability, or other fields needed for meaningful comparison.
02 / NORMALIZE
Make equivalent information comparable
Standardize fields such as location, compensation, and work arrangement so different sources can be evaluated on the same basis.
03 / VERIFY
Check whether the opportunity is still usable
Confirm that the listing is still active and that the underlying information is current enough to support a decision.
04 / DEDUPE
Count the opportunity once
Detect repeated postings across sources so the same role does not distort volume, scoring, or application tracking.
Garbage in, garbage out was a system-design problem—not a prompting problem.
Better evaluation depended on better inputs, so I improved the information before asking AI to interpret it.
V3 / EVALUATE
Turn “Do I want this?” into something the system could reason about.
Clean data made opportunities comparable, but comparison still required judgment.
I translated the criteria I was already using implicitly into an explicit 100-point evaluation framework. Instead of treating “fit” as one vague question, the system could reason across several dimensions—and I could see why one opportunity scored differently from another.
100-POINT EVALUATION FRAMEWORK
ROLE FIT
INTEREST
GROWTH
ZERO-TO-ONE
RESILIENCE
COMPENSATION
WORK ARRANGEMENT
DIMENSIONS WEIGHTED BASED ON RELATIVE IMPORTANCE
TWO DIFFERENT QUESTIONS
CAN I DO THE JOB?
Does the work itself align with my skills, interests, strengths, and the kind of problems I want to solve?
WILL I LOOK CREDIBLE ON PAPER?
Does my documented experience make me credible for screening against the requirements as written?
BEFORE
"Does this seem like a good job?"
Judgment was real, but the criteria were mostly implicit.
AFTER
"Why is this opportunity strong or weak?"
The system evaluated explicit dimensions that I could inspect, challenge, and revise.
AI became more useful once my own decision logic became explicit.
The framework gave the model something more useful than a vague instruction to “find good jobs”—but the score was still a recommendation, not a decision.
V4 / CALIBRATE
When the system disagreed with me, that became useful data.
Once the evaluation framework was explicit enough, I began having AI score opportunities against it.
I still reviewed the recommendations myself. When the system’s judgment and mine diverged, I treated the disagreement as a signal that either the criteria, the weighting, or the interpretation needed work—not as a reason to automatically accept one answer over the other.
AI EVALUATION
HUMAN REVIEW
AGREE?
KEEP FRAMEWORK
YES
NO
REFINE CRITERIA
DIAGNOSE WHY
DISAGREEMENT → DIAGNOSIS → REFINEMENT
BEFORE
AI score = answer
A recommendation could appear precise without revealing whether the underlying interpretation matched my actual judgment.
AFTER
AI score = hypothesis
The score became something I could inspect, challenge, and use to improve the framework.
LATER SIGNAL
The system could also learn from what actually generated interviews.
I later added patterns from applications that resulted in interviews as another prioritization signal. That signal could influence ranking without overriding the underlying evaluation framework.
The goal wasn’t to teach the system to agree with me. It was to make disagreement explainable.
Calibration turned the score from a black-box recommendation into an evolving decision system I could inspect and improve.
V5 / APPLICATION WORKFLOW
Once I knew which roles were worth pursuing, tailoring became the next bottleneck.
Application tailoring evolved the same way as the rest of the system: manually first, then with AI assistance, and finally as a codified workflow.
I started by editing every résumé and cover letter myself. Then I used AI conversationally to compare job descriptions against a detailed bank of my documented experience, accomplishments, projects, and metrics. Once that process became repeatable, I automated the mechanical parts while keeping claim validation, positioning, editing, and final approval with me.
01 / MANUAL
I tailored everything myself
JOB DESCRIPTION
READ + INTERPRET
EDIT RESUME
WRITE COVER LETTER
~20 – 30 min each
02 / AI-ASSISTED
AI helped me decide what to emphasize
JOB DESCRIPTION
AI ANALYSIS
RECOMMENDED POSITIONING
MANUAL EDITS
~10 – 15 min each
03 / WORKFLOW
The repeatable steps became a system
JOB DESCRIPTION
ANALYZE
REQUIREMENTS
SELECT BEST
BASE RESUME
TAILOR + DRAFT
HUMAN REVIEW + APPROVAL
~3 – 5 min human review
SYSTEM HANDLES
· Analyze full job description
· Identify central requirements and themes
· Select the closest existing résumé
· Create a role-specific copy
· Match documented accomplishments to the role
· Draft tailored résumé content
· Generate a complementary cover letter
· Maintain naming / formatting conventions
· Link materials back to the tracker
HUMAN RETAINS
· Whether to apply
· Whether the positioning is accurate
· Whether a claim is supported by real experience
· Final wording and messaging
· Résumé / cover-letter approval
· Final submission
~20 – 30 min → ~3 – 5 min
Hands-on tailoring time per application package after the workflow was codified.
The system could draft the materials. I still owned the claims and the decision.
Automation compressed the repetitive work without transferring accountability for what the application said—or whether it should be submitted.
AI + HUMAN
Automate the processing. Preserve the judgment.
A central design principle throughout the system was deciding where automation created leverage—and where human judgment should remain authoritative.
I used AI for repetitive information processing, structured comparison, requirement matching, and draft creation. I kept control over consequential decisions: whether a role was genuinely interesting, whether the recommendation made sense, whether my experience supported a claim, how I wanted to position myself, and whether anything should actually be submitted.
SYSTEM HANDLES
HUMAN OWNS
INGEST + STRUCTURE
Extraction · normalization · enrichment · deduplication
INTEREST
Do I actually want this?
SCREEN
Closed / duplicate / low-relevance opportunity detection
AMBIGUITY
Does the recommendation make sense?
EVALUATE
Preliminary scoring · requirement matching
CREDIBILITY
Does my documented experience support the claim?
PRIORITIZE
Surface stronger opportunities for review
TRADEOFFS
Compensation · growth · work arrangement · priorities
DRAFT
Resume positioning · cover-letter generation
POSITIONING
How do I want to tell the story?
CHECK
Formatting · naming · consistency
ACCOUNTABILITY
Final documents · whether and when to apply
The system can recommend. It cannot decide what I value.
The point was not autonomous job searching. It was moving my time away from repetitive processing and toward higher-value judgment.
PROCESS → FILTER → RECOMMEND → HUMAN DECISION → HUMAN APPROVAL
The strongest automation wasn’t where AI did the most. It was where responsibility was clearest.
Clear boundaries made the system faster without making it less accountable.
RESULTS
The system scaled the search without scaling the manual work.
The finished workflow could process hundreds of incoming recommendations, reduce duplicate and low-value review work, surface stronger opportunities for deeper attention, and generate tailored application materials for the roles I chose to pursue.
The biggest result was not simply faster processing. It was changing where I spent my time—from collecting, cleaning, comparing, and reformatting information toward evaluating opportunities, researching companies, preparing for interviews, and exercising judgment.
01 / SCALE
Hundreds of recommendations became manageable
The system now processes roughly 300 incoming recommendations per week and has evaluated about 1,500 unique opportunities across LinkedIn, Indeed, and BuiltIn.
02 / FASTER EVALUATION
Comparable review time fell by about 80%
Reviewing and scoring 20 opportunities went from roughly 100 minutes manually to about 20 minutes once the system could filter, enrich, deduplicate, and prioritize them first.
03 / FASTER APPLICATION PREP
Tailoring became a review task instead of a production task
Application materials that could take roughly 20–30 minutes each to tailor manually could be generated in seconds, leaving about 3–5 minutes of hands-on review and editing.
04 / MORE HIGH-VALUE TIME
Human effort moved toward work that benefited from judgment
The workflow freed more time for deeper evaluation, company research, interview preparation, and decision-making rather than repetitive information handling and document formatting.
JOB REVIEW
~100 min → ~20 min
per 20 opportunities
APPLICATION MATERIALS
~20 – 30 min → ~3 – 5 min
hands-on tailoring / review per application package
~300 RECOMMENDATIONS / WEEK · 2,000+ UNIQUE OPPORTUNITIES · ~120–150 TAILORED PACKAGES
The system didn’t make the search automatic. It made my attention more selective.
The value came from spending less time processing every opportunity equally—and more time on the opportunities that deserved deeper judgment.
LESSONS
What building this changed about how I think about AI systems.
01 / START WITH THE WORK
Good automation usually starts manually.
Performing each stage myself first helped me see the decisions embedded in the workflow, separate judgment from repetition, and formalize recurring patterns before automating them.
03 / MOVE HUMANS TOWARD JUDGMENT
The goal is not to remove the human. It is to improve where human effort goes.
The most useful automation moved my time away from repetitive processing and toward higher-value evaluation, interpretation, and decision-making.
02 / LET USAGE REVEAL THE NEXT PROBLEM
Real usage is a better roadmap than prediction.
Each improvement exposed the next constraint: ingestion revealed enrichment, enrichment revealed evaluation, evaluation made scoring possible, scoring created the need for calibration, and faster evaluation exposed application materials as the next bottleneck.
04 / MAKE RESPONSIBILITY EXPLICIT
The best AI systems have clear boundaries of authority.
The strongest parts of the system were not necessarily where AI did the most work. They were where the division of responsibility between automation and human judgment was clearest.
The point of AI wasn’t to make the system autonomous. It was to make the human work more valuable.
AI ENABLEMENT · HUMAN-IN-THE-LOOP · DECISION DESIGN · ITERATIVE SYSTEMS
NEXT PROJECT
Weekly Accuracy System
Turning QA data into a map of where to look next.