AI · Hospitality · Contract Intelligence · Stop-Sales Automation

Nothing Is Assumed: AI that reads hotel contracts
and refuses to guess.

A travel DMC was spending hours on every hotel contract and chasing stop-sale updates across dozens of formats. Accuavail is the Java and Python platform we built, powered by large language models, where AI reads, code verifies, and people approve. Every number on the output can be traced back to the page and clause it came from.

85%faster contract review
99%+verified extraction accuracy
23deterministic anomaly checks
0values without a source reference
ClientConfidential DMC / hospitality operator
IndustryTravel & Hospitality · Contracting · Inventory
Core stackJava, Spring Boot, Python, LLMs (Claude), PostgreSQL
EngagementPlatform build & ongoing operation
The challenge
In hospitality, one wrong guess shows up at the front desk.

Hundreds of hotel contracts landed in one inbox, and no two looked alike. They ran 5 to 30 pages each. Some read like spreadsheets, some like legal essays, and some were crooked scans. A single multi-property contract took 4-6 hours to read by hand. Meanwhile, stop-sale calendars kept arriving as PDFs, converted spreadsheets and email digests, each hotel with its own idea of what "closed" looks like.

The cost of slow

Contracting teams spent their days copying rates instead of negotiating them.

The cost of wrong

A misread season or missed stop-sale meant an oversold room and a guest already at check-in.

The trap of generic AI

Off-the-shelf tools filled gaps with confident, plausible answers. Plausible isn't true.

The format chaos

Digital PDFs, scans, Excel conversions and nested Outlook attachments, all in the same week.

Seven ways a stop-sale calendar lies

"Read a calendar, close a room" sounds simple. It isn't.

These are real failure modes we hit in production, and each one is now a named, hard-coded check:

01
A beige "on request" fill looked exactly like an empty cell.
02
A 7-room, 15-day stop-sale window hid inside one property's sheet.
03
Two legend entries shared one colour, so colour alone couldn't tell them apart.
04
Theme-indexed fills had no real colour value and silently dropped rows.
05
Files arrived with no property name, no date, or no legend.
06
A currency rate sat in a status cell where a legend code belonged.
07
Low-resolution scans where status truly couldn't be confirmed went to a human, not a guess.
Not a smarter prompt, but a program that measures a hex value and a pixel position, and can't be talked out of what it finds.
How it works

Two pipelines. One rule: AI reads, code verifies, people approve.

Contract Intelligence: per-document flow
01 / INGEST
Read the real layoutDigital PDFs parsed with structure intact; scans go through OCR and are flagged for reviewPDFBox · pdfjs · Tesseract
02 / EXTRACT
Zero-History AIThe LLM pulls every field into the schema, each with a page & clause referenceClaude · Python
03 / GATE
Go / no-go23 anomaly checks run, and the process stops 3 times for a human decisionJava · Spring Boot
04 / DELIVER
Verified outputStored, rendered to Excel, and recalculated in real Excel before releasePostgreSQL · POI · Excel COM
Stop-Sales Automation: per-calendar flow
01 / COLLECT
Unpack everythingOutlook digests, nested attachments and converted sheets, with signature images filtered outPython · Outlook parsing
02 / DECODE
Pixel-exact legendExact hex colour and rectangle position read from the source PDF, not the converterpdfjs · PDFBox
03 / CHECK
Named rule checksDuplicate swatches, missing legends and stray rates caught by rule, not by judgmentJava · pandas
04 / APPLY
Update or routeClean holds applied; uncertain cells sent to manual review with evidence attachedopenpyxl · ExcelJS
Architecture & tech stack

The right tool for each job, on purpose.

Reasoning-heavy work runs on AI. Anything that can be measured runs on deterministic code. This split is what lets the platform scale without letting errors in.

LAYER 1

Ingestion

Every format handled by its own parser, not one generic parser stretched to cover all of them.

Apache PDFBoxpdfjs-distTesseract OCROutlook email parsing
LAYER 2

AI reasoning

LLMs read contract language, reconcile the schema and judge edge cases. A second model gives an independent opinion.

ClaudeChatGPT (independent review)LLM document understanding
LAYER 3

Core services

The batch-processing engine, review application and rule engine with its 54 failure rules.

JavaSpring BootPythonNode.js
LAYER 4

Data & output

A structured, source-traced contract store and clean Excel outputs for operations teams.

PostgreSQLpandasopenpyxlApache POIExcelJS
LAYER 5

Verification

Every workbook is opened in real Excel and force-recalculated, because a library reading its own formula back isn't proof.

PowerShellExcel COM automationHuman sign-off
Our solution

Guardrails built from real mistakes, not theory.

// ZERO-HISTORY RULE

No memory, no assumptions

Before each extraction the AI states it has no prior knowledge of the property. Every value it returns must cite a page and clause, or it isn't returned.

// HUMAN-IN-THE-LOOP

3 go / no-go gates

The process pauses for a human decision at three points, so it never moves on just because the model feels confident.

// DETERMINISTIC CHECKS

23 anomaly checks

Conflicting validity periods, impossible rates and missing seasons are caught by Java rules that can't be talked past.

// PIXEL-EXACT

Legend verification

Stop-sale colours are matched on exact hex values and coordinates from the source PDF, never on a converter's fill colour.

// INSTITUTIONAL MEMORY

54 numbered failure rules

Every defect found in operation became a permanent, numbered rule, so the same mistake never comes back silently.

// MARKET-TESTED SCHEMA

Benchmarked against 50 DMCs

The contract schema was checked against how 50 real DMCs take in contracts, then audited twice by independent reviews before it shipped.

Results achieved

Faster, and more accurate at the same time.

85%
Faster contract review
99%+
Verified field-level accuracy
<15 min
Stop-sale update per property
100%
Values traced to source
Before
  • 4-6 hours per multi-property contract
  • Stop-sales keyed in by hand from mixed formats
  • Errors found at check-in
  • No audit trail for extracted values
After
  • Under 45 minutes, including human review
  • 99.5% of calendar cells matched to source automatically
  • Anomalies caught before data leaves the platform
  • Every value linked to its page & clause

[Client quote - add an approved testimonial here, or remove this block.]

Head of Contracting, Confidential DMC

Your contracts have the answers. We make sure AI doesn't invent them.

If your contracting, inventory or operations team is still reading PDFs line by line, let's talk about automating it without giving up accuracy.

Let's talk about your platform.

Whether it's a new Spring Boot build, a legacy modernisation, or something in between - tell us what you're working on and we'll respond within 24 hours.

✉️
Email
📍
Location
Jaipur, India · Global delivery
We reply within 24 hours.
✓ Thanks - we've received your message and will be in touch shortly.