ARIA — Enterprise AI Operations Intelligence
An LLM-assisted operational analyst for district and regional management workflows
- Role
- Designer, developer and tester of the workflow, business rules, instructions and validation cases
- Links
- Source is private or sanitized; no public repository is linked.
- Reading mode
At a glance
- Problem
- District managers interpreted multi-property reservation reports by hand, so follow-up priority and contact-rate judgments varied by property and by day.
- Architecture choice
- Validation and deterministic business rules run first; the language model only receives validated, aggregated, de-identified tables and writes narrative. Every number has a rule behind it.
- My contribution
- Defined the metric formulas, wrote the instructions and workflow configuration, designed prioritization and the output template, built the synthetic validation set and troubleshot model and gateway behavior.
- Verified so far
- Initial reservation-report use case built and tested with synthetic validation cases; metric output stable across repeated runs on the same report.
- Still open
- Scope is one use case. Narrative quality still needs human review and no business-impact figures are claimed.
Project overview
ARIA is a custom AI-based operational assistant developed for district and regional management workflows within an enterprise internal AI environment (Public Storage). The initial use case analyzes structured reservation reports across multiple properties: it validates the input, applies documented business logic, calculates standardized operational metrics, flags exceptions, prioritizes reservations that need follow-up and generates property-level recommendations with a human-readable executive summary. Deterministic rules own every number; the language model is scoped to interpretation, explanation and summary writing.
Problem and context
District managers receive multi-property reservation reports that are time-consuming to interpret consistently. Follow-up priorities, contact-rate expectations and same-day urgency were being judged by hand, which produced inconsistent decisions across properties and days.
The assistant runs inside an enterprise AI environment. All public material about it uses synthetic data; no customer records, internal prompts, gateway details or proprietary documents are reproduced in this portfolio.
Objectives
- Turn a raw reservation export into a standardized district summary in one pass
- Encode follow-up rules once so every property is evaluated the same way
- Prioritize reservations by urgency and conversion risk rather than by row order
- Produce recommendations a manager can act on without re-reading the source report
- Keep sensitive customer data out of model outputs
- Make output quality testable with repeatable validation cases
My contribution
- Defined the metric vocabulary (contact rate, same-day share, rental-over-phone completion) and the exact formulas behind each
- Wrote and iterated the model instructions and workflow configuration, separating deterministic calculation from interpretation
- Designed the prioritization rules and the output template for district, property and reservation-level sections
- Built the validation set: synthetic reports with known expected metrics and edge cases (missing fields, duplicates, inconsistent statuses)
- Troubleshot model, gateway and configuration behavior when outputs drifted or calls failed
Technology stack
- LLM workflow design
- Prompt & instruction engineering
- Structured output design
- Spreadsheet ingestion
- Business-rule modeling
- KPI calculation
- Regression test cases
- PII suppression
Results
- Initial use case built and tested: reservation reports are converted into a standardized summary, prioritized queue and recommendations
- Validation cases catch malformed inputs before interpretation
- Metric output is stable across repeated runs on the same report
Limitations
- Scope is the initial reservation-report use case; broader rollout and later phases are not assumed complete
- Narrative sections depend on model behavior and still require human review
- No quantitative business-impact figures are claimed
- The portfolio demonstration is a synthetic, client-side reproduction of the workflow concepts, not a connection to any enterprise system
Interactive demonstration
LLM interpretation versus deterministic calculation
Deterministic rules own
Validation, thresholds, every KPI, the priority score and the ordering of the queue. These are reproducible by hand.LLM-assisted interpretation owns
Plain-language explanation of calculated results, recommendation wording and the executive narrative. It references numbers; it never computes them. This demo uses fixed templates in its place.Typical volume, healthy contact rates, one property slightly behind.
Workflow
- Sample Report48 rows, state pass
- Data Validation0 issues, state pass
- Business Rulesthresholds, state pass
- Metric Calculationdeterministic, state pass
- Prioritization8 queued, state pass
- Dashboardrecommendations, state action
Stage 6: the dashboard, queue and recommendations. In the production workflow, the LLM writes the narrative from these calculated tables; it never recomputes them.
Reservations
48
4 properties
Contact rate
83%
40 contacted · target 75%
Same-day
5
10% of volume
Rental over phone
68%
17 of 25 phone reservations
Converted
48%
23 converted · 5 stalled
Contact rate by property
Bars show contacted ÷ total per property; the dashed reference is the 75% target.
Executive summary
Generated from the calculated tables by fixed templates in this demo.
- 48 valid reservations across 4 properties; 40 contacted (83%).
- 5 same-day move-ins (10%); 5 of them are in the top of the follow-up queue.
- Rental-over-phone completion 68% on 25 phone reservations.
- No properties need immediate action; 1 on watch.
- Input passed validation with no issues.
Prioritized follow-up queue
Top 8 open or stalled reservations by score.
P-101Northgate (synthetic)
ok17 res · contact 76% · same-day 1 · ROP 60% · stalled 2
- Metrics are within target ranges. Maintain the current contact cadence.
P-102Riverside (synthetic)
ok6 res · contact 100% · same-day 1 · ROP 100% · stalled 0
- Metrics are within target ranges. Maintain the current contact cadence.
P-103Harbor View (synthetic)
ok9 res · contact 89% · same-day 0 · ROP 100% · stalled 1
- Metrics are within target ranges. Maintain the current contact cadence.
P-104Summit Park (synthetic)
watch16 res · contact 81% · same-day 3 · ROP 56% · stalled 2
- Rental-over-phone completion 56% is under target (60%). Coach the phone close on the next team huddle.
contactRate = contacted / total sameDayShare = sameDayMoveIns / total rentalOverPhoneRate = completedOverPhone / phoneReservations (null when no phone reservations) conversionRate = converted / total priorityScore = 100·[sameDay] + 50·[attempts = 0] + min(ageHours, 72) + 20·[stalled] flag = action if contactRate < 0.55 flag = watch if contactRate < 0.75 or sameDayShare > 0.3 or ROP < 0.6 or stalled ≥ 4 rates suppressed when total < 5
Milestones
Initial ARIA use case built and tested
Reservation-report analysis workflow with validation cases.