Slides: AI-Driven Multi-Document Correlation for Financial Compliance - Varsha Shah, Independent
Source Video
AI-Driven Multi-Document Correlation for Financial Compliance - Varsha Shah, Independent
Relationship To World's Fair 2026
These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.
Related Scheduled Sessions
- No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck.
Extracted Slides
OCR text:
eco Sellen: | fy irate Thi ari
MA PTMe se ECR N Mae Cree nerceal's
correlation ror Enterprise
i Fa 7 is | aed . | o i vy ca a
are ae ae A ear cat ca a a ieee ah tie i aa
es eras CL Or eee eLes eens) UP eeer Moree eC Omran em sael
iP Cott ana - ya
1 Aeon son nO
A framework for cross-document fraud detection through relational intelligence, evaluated
across 3 million anonymized records and four jurisdictions. po
By Varsha Shah, Enterprise Technical Architect, USA _
‘ A,
OCR text:
reed aa ieee aaa _e 7 an a ;
at} Compliance Walp irom Osaraes Oalercsuan:.
ze Us Pa
Multi-Jurisdictional Growing Data Volumes Sophisticated Fraud
Tsay Enterprise financial records span ee
OX TUIarISTSHnLepe TAC et aerne LiseL payrol, tas, procurement, and Mocern fraud explons
regulations across global systems vransactions, ceneracng data ata ACONSIStENSeS Wat only surface
create bhad spots that manual and scate that overwhelns when compattng across related
rule-based tools cannot reconcile. (ofola zaan) (leuars eek a chy GOCUMeNt sets, Not within any
TnI Gl ma s1e1el£e)
Ruse based gnG NL? auger ves systems ooerauig atthe document eve ace structure yoercapag’e of
cetectg these cross-docament ancmales, ny
a
OCR text:
eo e
as seca es 2 eS eer eT aes
he itl Woy Docuinent-Level
: DS tA REET CRETE Stet ae
: rae eee Vi use ws arse ehcires oTeban
om a i + A
om - >
. Tradhuana, como ance tcals ova wate recerds mesolaten. Fraug that
“os (268 | 3,30
" 98 ait ELAM OLED CHET OLA ISTE SO(Gl LSM OT SDOv LATER MOTs b 220] SSS CT Chee Eee Lerma
I 2450 2.00 350 Be eee a Reo eee eke eee a Boece em eee
nterral vandation.
a
963° 2,00 3,77 880255}
388 14,15 17.60 2.50 11.50
e 663 1410 1695 5.70 10.80 4. The most costly fraud patterns are not found within a single
1210 1550 10,704, 60,50 ,
” document. They emerge in the space between document,
N | a
———————————~~_
he
OCR text:
— a eo. ; ony A
Lyte pap sete pears aS ae aT ere na ara a nee ae ee
or AB UCCGR Ole era CU ee hen eel Grnes UG Oh Ler mR Eeny)
CUR ee ee Grer a eCreeca rant,
Graph engine fies pg freanciad tc aids
PORN a Sere Rese SE
Foes Pale be ieee ek Oc reaSE| ESOS ESOT LO 16 OTO ICON Oa DEO
of AN fae oc ate.
rarerecer ts Pace ay ene ee eee
Eee eile ace oe ace ane em (Rea ca
Pe iene bee
Si ree errs etre cca Pe ee Oe LB ee Ep cee es a eS eo ee eee Gu eM a
Fae eee AR Oo Mays es eA aa a Oe Ro a ee | A FS
OCR text:
a oe
eee ‘ees [a oe enn aoe ere Peer is
Graph-Based Entity Correlation
A nm 2 eeeteaeete
ea n @: bey ne re eee eer ee ee
8 cent 4 Tue foundation of the frameworg os a Knowledde Graph tnat establishes
fas t+ @ :
Y ar ae Be aor eae rie Lee oo ao eC re cea aan ars Rs ria ere ee ce eee eral eae
y ee
aan . Fontes Morlideced eid “escled acioss Gatutdbleslate souhces
us
e Reationsiups surfaced between vencars, empoyces, accounts and tings
Couns rere One Rt] Doria le crema (<i c seu CcLeMeen ROLe Lea CLETIS IA NCe lo Le Olen tia aoa E
OCR text:
Porbatillty Distribution “s Sovon Ryst «
Se oe
@
Adaptive Probabilistic Risk Model . a
Ancina y siqnals from coreetec GoCuments gre agcregated onto Cal bratec at © oY ce a of
rok scores that proct ze Cases redur mG human review je - : ,
G , ag
CMe or mec omer eae mt ferscer ialer ee emerae nem ea asl es eee von ae eee eer | ad — a . =
Commer [ole icamrel @rerehermnoreiaM init rele bk aie © laa Mele] 6 Rit ea 9 aot e161 pi . "
Let ne ee
© Risk thresnolds configurable per purscicuon and business UAL Oe wecoies eaten Sonus
Okene wares onan enaew
Pinel Oh Swe Oem
ra
¢ >
OCR text:
r
my _ 7”
c A
N i
n : aan
\ . a ue 6
Ay oa + . $ rs
Na Ne - oo
1 : - Te. 8 CON \ nd
Evoaluotion Conditions a — a | ee
(on eee Deer Oneeuamae nen. N ; an
On bey . aan
- D| eas | eel
ain : - of
Dataset Scale Jurisdictions a _ a a
Approximately 3 million Four distnet regulatory a : . _
Lal Lt) ree
EIalerenaan lp 4-te Ma patei poate Raa rere ln ers aeRO E Asp} ecm Chat Orth 16 eas a = f A et oad
entes tall ; B >. | 0
“ee ca
mo a i ~ ™ ais
wa | Pa . So :
Time Horizon i ae
. \ : ~ Te
Five years Of historical Gata reflecting real-world enterprise a Ah’ SS ir
y on
cond:tions WA Pe ‘al
Sf bs
OCR text:
ao Pane a Cc rar
Ree ee el eer eee eee it .
Detection Performance Results
~91% ~87% ~0.89
oO O e
| ears tore! Recall F1 Score
Flagged cases couirmed as Go uine True fraud cases successful y surfaced Harpome ean balane 1g prec sar
rel prone ry Ears by the syste and recall
Perornance Was Negsured age nst dlapeed around trata cenved from conhrmec audit §id.ngqs cross
Pee ce) ym reel Le Gren ONC Ras UTR ner Mant Bel tate sn alC loka a
( ¥
OCR text:
CEC a cee ie ee eee ,
Operational Efficiency Gains
(0 0
ei ~40%
False Positive Rate Audit Workload Reduction
emma kee ei er ae ee er a ear eR ea ee aun le ME ar aa aS eae eg
are re ee ocr re rere AN
Be ype C aoa
OCR text:
Comparison Against Baseline Systents
High False Pete Rat ti Low False im
Positives ~38% 7 Positives ‘ . Positives ~9% ;
Document-Level <; Partial Cross- iy , Fuil Cross- fl
Only Be Document re Document :
Tow NekVoxahz-) Limited alii
No Adaptive | Document Continuous
Learning j Adaptability ; Adaptive Learning |
, | from Audit Fe- |
. pane a4 edback =,
C a
OCR text:
rroin Reactive to Fi Aran ooo er en sen mersserere
; Predictive
won aenre 7 ;
_—e ; Governance
Aru varna vars) .
PEt bs CEE eee
ay
\
Continuous
Monitoring
Pe eo oe ee ek ee i ae Me we Me ee een Mee ue See eae ale kal ‘
PS a ame ri a re er a ore ae Re See ec 4 a
OCR text:
— 1? os a o ot Cn 7a
ae ae oe a ae a | a et | aa eed ah aa omens Oye ESD TP eR ETON TTR
Practical De SACO PEC tas Came CLOTS R TCS se eee ces ete)
Es
Data Integration Jurisdictional Configuration
Ogden Maer nme 2) Neel Ene | CRN XG ees Ola Ba emo) CEN Gy Tye rormalization ayer must be oaramete gad fa
payroll PRP procurement anc tapcsystenis bxaisting irs la RES] OL tart OS ea To eee eo ME Ake Lbs OTe CR CSE nS IRAS!
Gdty pipe des Dun bo overaged were schemas Rone La OaD FORO SE Sd tea Len oS OTe OS Ard OL [Ome Da ean ad re Lo
Sia erates ee terre ne
Audit Team Alignment NTen UU tei ed
Invesugater work ows sheuld be recesgned to ears l ora exe nr oUt Es itera Ora als moles lot) be am Teceta ale tial
ES Tat se Lhs Lean S Gare CLACTON Ole] Cte maT eea Le Meats Se ICM Pee erm an Tec et oi cams Saar ee el etola mt ene ht CLE
Reronere kaa tthe lane IOLSL@ Lenten as 811A 16 NC ize Ie Lota tiberah sane Oe OL ai arenes mele ne gan eee va Om mate] arc Wel ame ae
Peeae HL RSTETESEL] [soe eee Ore Tne Oa bes i
/ ”
OCR text:
Thank you.
Varsha Shah — Enterprise Technical Architect
¢ :
Slide-Derived Subjects To Review
Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.