LIVE
GRAPH-GROUNDED AML INTELLIGENCE · TIGERGRAPH HACKATHON 2026

AML compliance,
graph-grounded.

An AML compliance copilot built on TigerGraph GraphRAG. Drop in any transaction, entity, or scenario — get a SAR-ready case file with OFAC sanctions check, multi-hop reasoning, and regulatory citations in 20 seconds. Benchmarked against vector RAG on 40 multi-hop questions.

+126%
ACCURACY vs BASIC RAG
−52%
TOKENS / CORRECT
+68%
BERT F1 vs P2
720×
FASTER THAN HUMAN
View on GitHub →
LIVE OFAC SCREEN LIVE
GRAPHRAG CORE — LIVE
84.6%
LLM-JUDGE PASS
3,197
AVG TOKENS / Q
16.25s
AVG LATENCY
RISK DISTRIBUTION LIVE
CO-PILOT PIPELINE
01
OFAC SDN Matcher
● 18,947 entities · RapidFuzz
02
Graph Retriever
● TigerGraph · 2-hop traversal
03
Risk Scorer
● HIGH / MED / LOW bands
04
SAR Generator
● Gemini 2.5 · cited draft
VS BASIC RAG
ACCURACY
+126%
TOKENS / Q
+9%
TOK / CORRECT
−52%
BERT F1 vs P2
+68%
✓ GRAPHRAG WINS — 40-Q BENCHMARK
0OFAC SDN ENTITIES INDEXED
3.4MAML CORPUS TOKENS
0BENCHMARK QUESTIONS GRADED
0× FASTER THAN HUMAN ANALYST
TRY IT YOURSELF

Audit any of the 40 questions.

Pick a benchmark question. Read all three pipeline answers side-by-side, with LLM-Judge pass/fail, BERTScore F1, token count, and latency for each. Click to expand judge reasoning. This is exactly the audit a compliance officer or hackathon judge would run.

Loading benchmark data...
THE BENCHMARK

Three pipelines.
One winner.

Same 40 hand-curated multi-hop AML questions. Same LLM (Gemini 2.5 Flash-Lite). Same data. Graded by LLM-as-Judge and BERTScore F1.

● VS BASIC RAG — GRAPHRAG WINS ON EVERY METRIC
+126% accuracy. −52% tokens. +68% cost.
+126% ACCURACY
+9% TOKENS / Q
−52% TOK / CORRECT
+68% BERT F1 vs P2
Pipeline Avg tokens Cost / query Latency LLM-Judge pass BERTScore F1 Tokens / correct
P1: LLM-only 748 $0.0002 4.71s 75.0% (30/40) 0.076 997
P2: Basic RAG 2,924 $0.0003 2.59s 37.5% (15/40) 0.187 7,799
P3: GraphRAG WINNER 3,197 $0.0029 16.25s 84.6% (33/39) 0.315 3,779
10 / 10
2-hop reasoning accuracy — the regime AML investigations live in. Basic RAG: 4 / 10.
2 / 4
Hybrid graph + text reasoning. The only pipeline that scores at all. Basic RAG: 0 / 5.
4.49×
Parallel-throughput speedup. Effective per-query latency drops to 3.62s under realistic concurrent load.
WHAT SANCTIONTRACE DOES

Four live modules.
One graph-grounded copilot.

Every transaction screened, every entity traced, every claim cited — from a TigerGraph knowledge graph, not from an LLM's parametric guess.

01
MODULE 01
Compliance Co-Pilot
Drop in a transaction, entity, or scenario. Get a SAR-ready case file in 20 seconds — OFAC match, risk score, regulatory citations, narrative draft.
→ CASE FILE GENERATED · 16.25s
02
MODULE 02
Knowledge Graph Explorer
Browse the live TigerGraph knowledge graph. Every node, every edge is queried in real time from a 3.4M-token AML corpus. Click any node to focus its neighborhood.
SEED SDN MATCH
→ ENTITY → CHUNK → ENTITY · 2-HOP
03
MODULE 03
Pipeline Comparison
Three pipelines, 40 hand-curated multi-hop AML questions, LLM-Judge + BERTScore evaluation. Watch where Basic RAG fails and GraphRAG succeeds.
P1
P2
P3 ★
→ 84.6% vs 37.5% · WINNER: GRAPHRAG
04
MODULE 04
Per-Question Explorer
Pick any of the 40 benchmark questions and see all three pipelines side-by-side, with LLM-Judge reasoning and BERTScore F1 surfaced per question. Audit anything.
P1 LLM-only · PASS · F1 0.289
P2 Basic RAG · PASS · F1 0.406
P3 GraphRAG · PASS · F1 0.567
★ JUDGE REASONING ATTACHED
HOW IT WORKS

Four stages.
One graph-grounded pipeline.

01
Ingest
FinCEN advisories, OFAC guidance, BSA regulations, Treasury enforcement actions — 3.4M tokens of source material chunked semantically and embedded.
02
Extract
TigerGraph GraphRAG extracts entities, relationships, and community summaries — turning unstructured text into a queryable knowledge graph.
03
Retrieve
For every question, the GraphRAG service performs hybrid vector + multi-hop graph traversal — Entity → DocumentChunk → Entity, up to 2–3 hops.
04
Generate
Gemini 2.5 Flash-Lite synthesizes the final answer using only the graph-grounded context, with regulatory citations baked into the response.
SRC
Source corpus
FinCEN · OFAC · BSA · Treasury
3.4M tok
TG
TigerGraph 4.2
Knowledge graph · vector index
● LIVE
GR
GraphRAG service
Hybrid retrieval · 2-hop traversal
:8000
LLM
Gemini 2.5 Flash-Lite
Cited synthesis · grounded output
● LIVE
SAR-READY CASE FILE
20 SECONDS · WITH CITATIONS
SAMPLE OUTPUT

What the Co-Pilot actually returns.

Two real case files generated by the Co-Pilot — same structure a compliance analyst receives in 20 seconds. Risk band, OFAC SDN match, regulatory citations, recommended actions, and an analyst-editable SAR narrative draft.

CASE-2026-0517-A
HIGH · 78/100
QUERY · ENTITY
BANCO NACIONAL DE CUBA
[OFAC] SDN Match
BANCO NACIONAL DE CUBA · 90% string similarity · Program: CUBA · Type: BANK
[RISK] Signals
  • Receiver appears on OFAC Cuba Sanctions Program SDN list
  • USD-denominated transfer to embargoed jurisdiction
  • Counterparty controlled by Government of Cuba
[CITATIONS]
31 CFR § 515 CACR § 515.207 OFAC Cuba Sanctions Program FinCEN SAR Filing Instructions
[ACTIONS]
  • Block transaction pre-settlement
  • File SAR within 30 days (FinCEN Form 111)
  • Escalate to MLRO; preserve full audit trail
SAR narrative draft (analyst-editable)
On 2026-05-17 the filing institution detected an outbound USD wire of $250,000 from account C12345678 to a counterparty identified as BANCO NACIONAL DE CUBA, a financial institution appearing on the U.S. Treasury OFAC Specially Designated Nationals (SDN) List under the Cuba Sanctions Program. The transaction was blocked pre-settlement pursuant to 31 CFR § 515.201. Given the receiver's status as a state-controlled bank of an embargoed jurisdiction, the institution assesses this activity as high-risk for facilitating prohibited financial dealings with the Government of Cuba.
Generated: 1,742 tokens · 18.4s · Gemini 2.5 Flash-Lite ● GRAPHRAG-GROUNDED
CASE-2026-0517-B
HIGH · 84/100
QUERY · SCENARIO
12 cash deposits between $8,500–$9,500 across 4 branches in 5 days, totaling $108,000. No source-of-funds documentation.
[OFAC] SDN Match
No SDN matches on subject; no counterparty named.
[RISK] Signals
  • Classic structuring pattern: all deposits below $10,000 CTR threshold
  • Multiple branches in short window — evasion of single-branch detection
  • Aggregate ($108K) far exceeds the threshold the individual deposits avoid
  • No source-of-funds documentation provided
[CITATIONS]
31 USC § 5324 31 CFR § 1010.314 FinCEN SAR Activity Review #19 BSA § 5318(g)
[ACTIONS]
  • File SAR within 30 days, code FIU.S (structuring)
  • Place customer under enhanced due diligence (EDD)
  • Cross-reference against related-party transactions for 90 days prior
SAR narrative draft (analyst-editable)
Between 2026-05-12 and 2026-05-16 the subject conducted 12 cash deposits, each in amounts between $8,500 and $9,500, across 4 separate branches of the filing institution. The deposits aggregated to $108,000 — a sum that would have triggered a Currency Transaction Report (CTR) if conducted in a single transaction. The pattern strongly suggests an attempt to evade the $10,000 reporting threshold of 31 CFR § 1010.311, in violation of the anti-structuring provisions of 31 USC § 5324. No legitimate source of funds was documented at any deposit. The institution recommends filing a Suspicious Activity Report and initiating enhanced due diligence.
Generated: 1,856 tokens · 19.1s · Gemini 2.5 Flash-Lite ● GRAPHRAG-GROUNDED
● Three more pre-cached cases live in the dashboard — open the demo to explore
// SANCTIONTRACE · v1.0 · TIGERGRAPH GRAPHRAG INFERENCE HACKATHON 2026
+126% accuracy.
−52% tokens per correct answer.
+68% cost per correct answer.

Open source.
Reproducible.
Graph-grounded.

Three pipelines benchmarked on the same 40 multi-hop AML questions. Code, data, evaluation harness — everything to reproduce the numbers above is in the repo.

ABOUT

What is SanctionTrace?

TIGERGRAPH GRAPHRAG INFERENCE HACKATHON · 2026

SanctionTrace is an AML (anti-money-laundering) compliance copilot that turns a 4-hour analyst workflow into a 20-second graph-grounded case file.

The problem

Compliance analysts spend hours manually pulling transactions, running OFAC screenings, walking counterparty relationships, and drafting Suspicious Activity Reports. Vector RAG copilots have tried to automate this and failed — vector search retrieves similar text, but it can't reason across the relationships AML investigations live in.

The solution

SanctionTrace builds a TigerGraph knowledge graph from a 3.4M-token AML corpus (FinCEN advisories, OFAC guidance, BSA regulations) plus 18,947 OFAC SDN entities, then puts a GraphRAG retrieval layer between Gemini 2.5 Flash-Lite and the analyst.

Results vs Basic RAG

  • +126% LLM-Judge pass rate (84.6% vs 37.5%)
  • −52% tokens per correct answer (3,779 vs 7,799)
  • +68% cost per correct answer ($0.0021 vs $0.0054)
  • +125% on 2-hop questions (9/10 vs 4/10) — the exact regime AML lives in
  • 2/5 vs 0/5 on hybrid graph + text questions
  • 720× faster than the 4-hour human-analyst baseline
STACK

What it's built on

REPRODUCIBLE · DOCKERIZED · OPEN-SOURCE

Graph layer

  • TigerGraph 4.2 Savanna — hosted graph database
  • TigerGraph GraphRAG service — Dockerized retrieval + extraction
  • Hybrid vector search + multi-hop graph traversal

LLM + embeddings

  • Gemini 2.5 Flash-Lite — synthesis + LLM-as-Judge
  • models/gemini-embedding-001 — 1,536-dimensional embeddings
  • ChromaDB — Basic RAG baseline (P2 only)

Sanctions screen

  • OFAC SDN — 18,947 specially-designated entities
  • RapidFuzz — string-similarity matching for entity resolution

Dashboard + landing

  • Streamlit + Plotly — interactive co-pilot dashboard
  • Vanilla SVG — live knowledge-graph rendering
  • Hand-rolled HTML / CSS — this landing page

Evaluation

  • 40 hand-curated multi-hop AML questions
  • LLM-as-Judge — PASS / FAIL grading
  • BERTScore F1 — semantic similarity to ground truth
  • Parallel-throughput measurement — 4.49× speedup
GET IN TOUCH

Contact

TIGERGRAPH GRAPHRAG INFERENCE HACKATHON · 2026

SanctionTrace was built solo by Akansha Mishra over seven days for the TigerGraph GraphRAG Inference Hackathon 2026. Reach out for questions, partnerships, or hiring conversations.

⚡ EMAIL
mishakansha.01@gmail.com
★ GITHUB
github.com/AMSecretAgent/sanctiontrace