Ramona AI — Virtual Knowledge Graph Observatory

A running Virtual Knowledge Graph (VKG) over a public-interest human-trafficking observatory, built on the ontology-based-data-access paradigm of Xiao, Ding, Cogrel and Calvanese (2019). Prepared for the Dialogue Residency at the Institute of Logic and Computation, TU Wien, May to July 2026.

What this repository contains

• Postgres 16 with Ramona's operational schema (10 tables: offers, recruiters, platforms, red flags, classifications, survivor patterns, trafficking routes, and provenance) plus the v1.0 observatory extension (victim-centric tables aligned to CTDC).
• RATR-O v1.0 (ontology/ratr-o.ttl): the Ramona Anti-Trafficking Recruitment Ontology in OWL 2 QL, aligned to the generalisable fragment of HUTRO (Human Trafficking Ontology, BioPortal, 2025) via skos:closeMatch.
• Ontop mapping (mappings/ramona.obda) connecting the Postgres schema to RATR-O.
• SPARQL endpoint served by Ontop on port 8080.
• Fourteen reference SPARQL queries (queries/Q1..Q14) covering Ramona operational questions and observatory statistics.
• A multi-source dataset pipeline under scripts/: press scraping (Mexico + LatAm), CTDC (IOM) ingestion, rule-based classification into Type A (classic red flags) and Type B (credible-lethal patterns) tracks.
• Evidence source catalogue (docs/EVIDENCE_SOURCES.md) and CTDC–HUTRO–RATR-O mapping workbook (docs/CTDC_HUTRO_RATRO_mapping.xlsx).

No personally identifiable information appears anywhere in the demo. All victim records are anonymised: CTDC records use Microsoft Research's differential-privacy synthetic generation, and press records are aggregated at the article level with outlet, title, URL and structured classification only.

The "nobody is exempt" framing

The observatory publishes two parallel classifications for every case:

Track	What it captures	Examples
Type A — classic red flags	The 7 red flags from Ramona's public infographic: high salary, flexible hours, no experience needed, advance payment, urgent acceptance, distant interview with paid travel, job in another state.	Obvious scam-style offers.
Type B — credible-lethal patterns	Plausible offers (market salary, realistic role) whose signal is logistic: no company logo, nameless WhatsApp, meet at a multi-use building, come alone, no ID, Messenger-to-WhatsApp funnel, personal-data filtering, atypical hours, last-minute relocation.	The Edith Guadalupe family of cases.

On measuring the Type-A vs Type-B distribution: an earlier draft reported ~10x more Type-B-only than Type-A-only cases from press. That figure was withdrawn after validation: press articles describe outcomes, not offer terms (an LLM read found only ~8% of articles contain the offer), so the ratio cannot be measured from press. See docs/CLASSIFIER_VALIDATION.md. The Type-A/Type-B distinction is a modelling contribution of the ontology; its empirical distribution will be measured from Ramona's operational reports, where the offer text exists. The press observatory's measurable contributions are geographic coverage (18 countries), volume, provenance, and a documented-case corpus.

Prerequisites

• Docker Desktop (or any Docker engine) with Docker Compose v2.
• About 1 GB of free disk and 2 GB of free memory.
• Python 3.11+ for the ingestion and scraping scripts.
• A browser for the Ontop SPARQL UI.

Start the VKG demo

From this directory:

docker compose up -d

First run:

1. pulls Postgres 16 and Ontop 5.2.0;
2. creates the ramona database and runs db/01_schema.sql, db/02_seed_reference.sql, db/03_seed_cases.sql, db/04_schema_extension_v1.sql via /docker-entrypoint-initdb.d/;
3. starts Ontop with ontology/ratr-o.ttl + mappings/ramona.obda.

Wait for Ontop to report as healthy:

docker compose logs -f ontop

Then open http://localhost:8080/ for the SPARQL UI, or run the whole reference query suite:

./scripts/test_queries.sh

Populate the observatory

# one-time
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# scrape press
python scripts/scrape_press.py --out data/raw/press.jsonl --max-per-query 100

# classify into Type A / Type B
python scripts/classify_press.py \
    --in data/raw/press.jsonl \
    --out data/processed/press_cases.jsonl \
    --summary data/processed/press_summary.json

# ingest press into Postgres
DSN="postgresql://ramona:ramona@localhost:5432/ramona" \
python scripts/ingest_press.py --in data/processed/press_cases.jsonl

# ingest CTDC (once you have the CSV from https://www.ctdatacollaborative.org/page/global-synthetic-dataset)
DSN="postgresql://ramona:ramona@localhost:5432/ramona" \
python scripts/ingest_ctdc.py --csv ~/Downloads/ctdc_global_synthetic_v2025.csv

Reference queries

File	What it answers
Q1_highrisk_by_platform.rq	High-risk and fraudulent offers grouped by platform (Ramona operational).
Q2_multicountry_recruiters.rq	Cross-border recruiters with multiple flags this quarter.
Q3_pattern_matches.rq	Offers matching survivor patterns, plus source that documents each pattern.
Q4_route_unclassified.rq	Offers on a known trafficking route that have not been classified.
Q5_cjng_salary_pattern.rq	Offers matching the CJNG post-Rancho Izaguirre pattern.
Q6_edith_pattern_bundle.rq	Offers bundling the Edith Guadalupe credible-lethal patterns.
Q7_messenger_whatsapp_funnel.rq	Messenger-to-WhatsApp funnel with personal-data filtering.
Q8_fraudulent_counts_by_country.rq	Confirmed-fraudulent offers by country, with average risk score.
Q9_ctdc_exploitation_by_country.rq	CTDC victims by country of exploitation and exploitation type.
Q10_type_b_only_cases.rq	Press cases that match Type B patterns but not Type A red flags.
Q11_recruitment_channels_distribution.rq	First-contact communication tool distribution across all sources.
Q12_cross_source_age_gender.rq	Age × gender × exploitation type (requires ≥3 victims per cell).
Q13_means_of_control_stack.rq	Means of control by exploitation type.
Q14_mexico_press_vs_ctdc.rq	Mexico: press-documented vs CTDC-registered victims (cross-source federation).

| Q15_fraudscheme_by_country.rq | Distribución de FraudSchemes por país (cross-source). | | Q16_palermo_triad_coverage.rq | Cobertura del triángulo Palermo (Action × Means × Purpose) por TIPEvent. | | Q17_scheme_vs_tool_heatmap.rq | Heatmap FraudScheme × CommunicationTool. | | Q18_age_gender_per_scheme.rq | Perfil demográfico (edad × género) por FraudScheme. | | Q19_outlets_by_scheme.rq | Outlets de prensa que más documentan cada esquema. | | Q20_type_b_over_a_by_country.rq | Ratio Type-B:Type-A por país. | | Q21_pattern_cooccurrence.rq | Pares de patrones que co-ocurren en el mismo artículo. | | Q22_recruitment_x_exploitation.rq | Método de reclutamiento × tipo de explotación (CTDC). | | Q23_whatsapp_funnel_schemes.rq | Esquemas que comparten WhatsApp como primer canal. | | Q24_new_routes_from_press.rq | Rutas nuevas documentadas por prensa aún no atestadas en CTDC. | | Q25_trend_by_year.rq | Tendencia 2002-2023 de víctimas registradas en CTDC. | | Q26_means_of_control_network.rq | Red de co-ocurrencia de Means of Control (CTDC). | | Q27_vulnerability_x_exploitation.rq | Factor de vulnerabilidad × tipo de explotación. | | Q28_crossborder_with_press.rq | Corredores transfronterizos con prensa reciente. | | Q29_edith_cohort.rq | Cohorte Edith Guadalupe (Type-B-only explícito). | | Q30_schemes_in_both_sources.rq | FraudSchemes atestados en CTDC y prensa (intersección). |

Layout

ramona_vkg_demo/
├── README.md
├── docker-compose.yml
├── requirements.txt
├── db/
│   ├── 01_schema.sql
│   ├── 02_seed_reference.sql
│   ├── 03_seed_cases.sql
│   └── 04_schema_extension_v1.sql          # victim-centric v1.0 extension
├── ontology/
│   └── ratr-o.ttl                          # RATR-O v1.0 in OWL 2 QL, aligned to HUTRO
├── mappings/
│   └── ramona.obda                         # Ontop native mapping, operational + observatory layers
├── ontop/
│   └── ontop.properties
├── queries/
│   ├── Q1..Q8_*.rq                         # Ramona operational reference queries
│   └── Q9..Q14_*.rq                        # Observatory-layer queries
├── scripts/
│   ├── scrape_press.py                     # Google News RSS + trafilatura
│   ├── classify_press.py                   # Type A / Type B rule-based classification
│   ├── ingest_press.py                     # press → Postgres observatory tables
│   ├── ingest_ctdc.py                      # CTDC CSV → Postgres observatory tables
│   └── test_queries.sh
├── data/
│   ├── raw/                                # scraper output (JSONL, git-ignored)
│   └── processed/                          # classified output (JSONL + summary JSON)
└── docs/
    ├── EVIDENCE_SOURCES.md                 # real-world sources behind every pattern and route
    └── CTDC_HUTRO_RATRO_mapping.xlsx       # mapping workbook prepared for Prof. Ortiz

Stopping and cleaning up

docker compose down        # stop containers, keep the volume
docker compose down -v     # stop and wipe the Postgres volume

Credits and licence

• Ontology (RATR-O) and scholarly outputs: CC BY 4.0.
• Demonstrator code (SQL, OBDA, scripts, README): Apache 2.0.
• CTDC Global Synthetic Dataset: IOM and partners, CC BY 4.0 with source attribution.
• HUTRO: Danial Zemchal Tesfahans, BioPortal, 2025.
• Framework: Xiao, G., Ding, L., Cogrel, B. and Calvanese, D. (2019). Virtual Knowledge Graphs: An Overview of Systems and Use Cases. Data Intelligence, 1(3), 201–223. DOI: 10.1162/dint_a_00011.
• Ontop: Calvanese et al. (2017), https://ontop-vkg.org.

Prepared by Daniela Camberos (Ramona AI) for the Dialogue Residency at TU Wien, Institute of Logic and Computation.