LLM Urban Data Pipeline: Catalog → Filter → Code → Retry (NYC Taxi + LEHD)

1 1. Introduction

*[To be completed]* Motivate the problem for R1 readers; state research questions; preview contributions (C1, C2, …); outline paper structure.

2 2. Related Work

*[To be completed]* Position against prior work (SSI, IoT identity, urban systems, agent ops). End with an explicit gap this paper addresses.

3 3. Background and Problem Statement

*[To be completed]* Domain context, definitions, and the concrete setting (e.g. urban–Web3, agent deployments, CEIN device class).

4 4. Methods and Research Design

*[To be completed]* Research design: data sources, protocols, metrics, ethics/consent, and how claims will be evaluated.

5 5. System Design and Architecture

*[To be completed]* Architecture, interfaces, threat model or operational assumptions, and what is implemented vs planned.

6 6. Evidence, Data, and Findings

• **Benchmark:** 400 queries (NYC Taxi + LEHD), plus unanswerable cases.
• **Full system:** **60.9%** accuracy vs **6.8%** for a naïve LLM-only baseline.
• **Largest gains:** explicit filtering step, dataset preview, and retry loop.
• **Graph-specific tasks** (community detection via Combo algorithm, centrality, density, PageRank) are supported but remain error-prone on multigraph and directionality details.

7 7. Discussion

Framing

The research ideas we pursue in iShareHow Labs are exactly what we are building with Agents and the large system upgrades now in production — and they require established data feeds. As the business grows, we collect more operational evidence on these concepts (Google Analytics, Python pipelines, MCP tools) while showing the power of LLMs in real work: closing the gap for urban communities, not only demoing models.

Editor-in-Chief note: this program treats decentralized and urban transactional systems (including Web3 substrates) as infrastructure for verifiable, optimizable layers that support algorithmic justice — not as a personal focus on trading, asset pricing, or institutional finance.

End-to-end pipeline (prompt chaining + chain-of-thought)

1. **Data-catalog selection** — LLM chooses the relevant catalog from the query.

2. **Dataset selection + preview** — metadata + column statistics.

3. **Filtering-condition generation** — LLM proposes filters using stats and docs.

4. **Code generation** — LLM produces executable Python with chain-of-thought.

5. **Execution + single retry** on failure.

Open extension surface

Add a CEIN identity attribute (or edge label) to the STTN node/arc schema, then measure how attribution, anomaly detection, community fairness, or optimization metrics change. Explainability can be layered on the resulting graph metrics or on the LLM-generated analysis code.

Labs surfaces to collect on

• Evidence Canvas + Seminar Calendar & Tracker (CUSP seminars, Urban Data Science Showcase, OSCUR workshops, UrbCom 2026).
• Connectivity GraphRAG (`labs_connectivity_*`) and journal workspace.
• Agents MCP / GA / Python execution loops as operational data feeds that feed scholarship.

Research identity

Methods center on machine learning, deep learning, data mining, optimization, network science and graph methods (GNNs, spatio-temporal transactional networks), explainable/interpretable AI, and privacy-preserving / on-device AI — applied to secure proximity-aware identity systems, complex networked infrastructures, and verifiable transactional layers for algorithmic justice.

STTN vs Neo4j Connectivity (peer alignment)

Peers define an STTN as:

> Directed, labeled multigraph that unifies spatial, temporal, and transactional attributes (nodes, arcs, space/time labels, and attribute functions). Converted at runtime into Pandas / GeoPandas / NetworkX objects for analysis.

What we already built with Neo4j (Connectivity / GraphRAG) is the index + provenance layer over real Labs/Agents/Web3 systems — same scientific direction, not yet the full peer algebra.

Bridge now live: `labsconnectivitysttn` projects Connectivity into that STTN object and converts at runtime to Pandas node/arc frames + NetworkX metrics (density, PageRank, centrality), with CEIN identity annotation via `labsconnectivitysttn_cein`.

Still thin vs peers (honest): dense CRS/geometry, interval space–time labels on most nodes/arcs, first-class transactional `tx_*` enrichment, Combo community detection, and CEIN coverage until annotated. GraphRAG answers questions; STTN export is what peers publish analysis against.

8 8. Limitations and Threats to Validity

*[To be completed]* Scope limits, sample bias, measurement error, generalizability, and what would falsify the claims.

9 9. Conclusion and Future Work

*[To be completed]* Restate contributions; future work; path to peer review / venue.

10 Acknowledgments

*[Optional]* Funding, collaborators, community partners.

11 References

*[To be completed]* Use a consistent style (IEEE or APA). Prefer DOIs. List only sources actually cited in the body.