Integrating Gemini for Natural Language Analytics in IoT Fleet Management

cox_2m · Tech · 2024 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Self-service analytics for IoT fleet-management data

Cox 2M had a lean analytics team serving high-volume internal ad hoc requests across large IoT datasets. Single business questions could take up to a week to answer, making insights stale and keeping analysts away from higher-value customer-facing work. The production goal was to let non-technical users ask data questions in natural language, understand KPI changes, and correct semantic misunderstandings without leaving the analytics workflow.

Constraints:

• Cox 2M operates automotive lot management, small business fleet management, and industrial/supply-chain IoT use cases
• the analytics team was lean and could not scale a high-touch cost-to-serve model
• business users needed useful answers before a week-old insight became irrelevant
• analytics had to run on live cloud data platforms rather than static reports
• semantic definitions had to remain consistent across ThoughtSpot and Looker-modeled data
• Gemini-generated natural language had to be auditable by business users through tokens and feedback
• vendor-reported metrics came from a Google Cloud Next customer presentation rather than an independent study

Architecture

Cox 2M used ThoughtSpot on Google Cloud, connected to Google BigQuery and other Google data services, with Gemini powering natural-language analytics features. ThoughtSpot Liveboards and Sage let business users ask questions, drill into KPI changes, and receive natural-language explanations. Gemini powers three main capabilities: change-analysis narratives, full-liveboard summarization, and Sage natural-language query. The system translates business-language questions into SQL and visualizations against live data, exposes tokens so users can validate how terms mapped to columns, and captures user feedback as reusable query fragments for future questions.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  Cox_2M_IoT_operations_data["Cox 2M IoT operations data"] --> Google_data_platform_layer["Google data platform layer"]
  Google_data_platform_layer --> Looker_Modelo_semantic_layer["Looker Modelo semantic layer"]
  Google_data_platform_layer --> ThoughtSpot_live_query_engine["ThoughtSpot live query engine"]
  Looker_Modelo_semantic_layer --> ThoughtSpot_live_query_engine
  ThoughtSpot_live_query_engine --> ThoughtSpot_Liveboards["ThoughtSpot Liveboards"]
  ThoughtSpot_Liveboards --> Change_analysis["Change analysis"]
  ThoughtSpot_Liveboards --> Liveboard_summarization["Liveboard summarization"]
  Business_user_interface["Business user interface"] --> Sage_natural_language_query["Sage natural-language query"]
  Sage_natural_language_query --> Gemini_LLM["Gemini LLM"]
  Gemini_LLM --> Change_analysis
  Gemini_LLM --> Liveboard_summarization
  Sage_natural_language_query --> ThoughtSpot_live_query_engine
  Sage_natural_language_query --> Token_validation_UI["Token validation UI"]
  Token_validation_UI --> Feedback_correction_interface["Feedback correction interface"]
  Feedback_correction_interface --> Feedback_fragment_store["Feedback fragment store"]
  Feedback_fragment_store --> Sage_natural_language_query
  ThoughtSpot_Liveboards --> Business_user_interface

Components

Data flow

IoT and business data sit in Google-connected data platforms such as BigQuery, Cloud SQL, AlloyDB, and Looker Modelo semantic definitions. ThoughtSpot hosts on Google Cloud and issues live queries over those platforms. Users view Liveboards, trigger KPI change analysis, ask Sage natural-language questions, and receive SQL-backed visualizations. Gemini generates natural-language narrations, summarizes liveboards, explains change contributors, and helps translate natural-language questions into SQL. When the system misunderstands a term such as popular plans, the user edits tokenized fragments; those fragments are stored for subsequent queries and reusable classes of problems.

Why this architecture

The architecture connects natural-language interaction directly to live governed data because Cox 2M's bottleneck was not a lack of dashboards but a lack of timely, self-service answers. ThoughtSpot provides the analytics interface, live queries, liveboards, drilldowns, and semantic integration; Gemini supplies language translation, narration, and summarization; Google data services provide live data access. The token-validation and feedback-fragment loop keeps the interface usable when business terms do not map cleanly to columns.

Models used: Gemini · ThoughtSpot Sage LLM capabilities

Scale

• Cox 2M reported 11.4 billion assets tracked over its lifespan.
  “11.4 billion assets tracked in our lifespan where we currently measure you know over 29,000 miles of asset movements every hour”transcript, stated
• Cox 2M reported over 29,000 miles of asset movement measured every hour.
  “over 29,000 miles of asset movements every hour”transcript, stated
• Ad hoc analytics requests previously took upwards of a week.
  “it took us you know upwards of a week to produce a single ad hoc request for our business”transcript, stated
• After the change, Cox 2M reported less than an hour on ad hoc requests when they handled them at all.
  “if we even spend any time at all on ad hoc requests which is very little we spend less than an hour reduced time to insights by 88%”transcript, stated
• The reported time-to-insight reduction was 88%.
  “reduced time to insights by 88%”transcript, stated
• The demo analyzed a KPI increase of more than 100% in roaming call minutes.
  “my roaming call minutes actually has gone up by 100 more than 100%”transcript, stated

Alternatives considered (roads not taken)

Continue serving ad hoc analytics through the lean internal analytics team

Rejected because: The model was not scalable for a small team and produced insights too late to remain relevant.

Traditional BI self-service for analytics specialists only

Rejected because: Cox 2M needed self-service for business users, not just BI and analytics teams.

Require users to inspect every dashboard chart manually

Rejected because: Gemini summarization reduces cognitive load by highlighting expected and unexpected liveboard changes.

Treat LLM query translation as opaque

Rejected because: ThoughtSpot exposes tokens and feedback editing so users can validate and correct mappings.

Patterns

query-rewriting (retrieval)

Sage translates natural-language business questions into SQL-backed visualizations over live data.

human-in-the-loop-gate (workflow)

Business users validate token mappings and submit corrections when a natural-language term maps to the wrong query semantics.

data-quality-loop (data)

Corrections are stored as smaller fragments and reused for later queries and related problem classes.

structured-output-contract (interface)

Sage exposes tokenized query fragments and column-name mappings that users can inspect and edit instead of treating the LLM output as free-form prose.

rag-retrieval (retrieval)

Natural-language questions execute as live queries over governed data platforms rather than relying on static dashboard extracts.

tool-use-orchestration (workflow)

Gemini powers narrative and query features while ThoughtSpot orchestrates liveboards, change analysis, SQL execution, visualizations, and feedback capture.

Tradeoffs

Move from analyst-served requests to business-user natural language analytics.

Chose ThoughtSpot Sage with Gemini-powered natural-language querying over continuing to route most ad hoc requests through the lean analytics team.

Rationale: The old model was slow and not scalable; natural-language self-service freed the analytics team to focus on high-value customer outcomes.

Run natural-language analytics against live data platforms.

Chose live BigQuery and Google data-platform queries through ThoughtSpot over pre-aggregated or static dashboard-only analytics.

Rationale: Live query integration let users ask natural-language questions, drill down, and run change analysis on current governed data.

Expose token-level validation and correction.

Chose user-visible tokens plus feedback correction over opaque natural-language-to-SQL generation.

Rationale: Token validation lets users catch semantic mismatches and train reusable fragments for later queries.

Use Gemini for narrative summarization instead of chart-by-chart interpretation.

Chose Gemini-generated change narratives and liveboard summaries over manual interpretation of every KPI chart and dashboard visualization.

Rationale: Narratives translate data-speak into business language and help users focus on what changed.

Gotchas (what bit them)

Analytics self-service was not truly reaching business users.

Root cause: Pre-generative-AI self-service still mostly served BI and analytics teams rather than end users.

Resolution: ThoughtSpot Sage put a natural-language interface in front of business users so they could ask questions directly.

Week-long ad hoc analytics made insights stale.

Root cause: A lean analytics team spent too much time serving internal one-off requests.

Resolution: Cox 2M centralized analytics and adopted ThoughtSpot/Google/Gemini capabilities to reduce time to insight.

Natural-language terms may not map to dataset terms.

Root cause: Users can ask for concepts such as most popular plans that do not exist verbatim in the dataset.

Resolution: The system maps terms into tokens, lets users correct the query, and stores fragments for reuse.

Dashboard volume can hide the actual business change.

Root cause: Users may need to inspect many charts individually to understand what changed.

Resolution: Gemini summarizes liveboards and highlights expected and unexpected changes.

LLM-generated answers still need transparent validation.

Root cause: Business users need confidence that the natural-language question was translated to the correct columns and SQL.

Resolution: The UI exposes tokens as column names so users can inspect whether Gemini did the work correctly.

Evaluation

The presentation evaluates the system through customer-reported operational metrics and live demo behavior rather than a controlled benchmark. It reports reduced time to insight, shows business-user query examples, demonstrates change-analysis narratives, token validation, complex query translation to SQL and visualization, and feedback capture for future queries.

Metrics: 88% reduction in time to insight · ad hoc request time from upwards of a week to less than an hour · 11.4 billion assets tracked · over 29,000 miles of asset movement measured every hour · 100%+ KPI increase analyzed in demo · top-three customer query translated to SQL and visualization · feedback fragments stored for subsequent queries

Outcomes

• Cox 2M reported an 88% reduction in time to insight.
• Ad hoc request handling dropped from upwards of a week to less than an hour when such requests still occurred.
• Non-technical users can ask natural-language questions, receive visualizations, and validate token mappings.
• Gemini powers natural-language narrations, change analysis, liveboard summarization, and Sage natural-language query.
• Corrected query fragments are stored for reuse across subsequent queries and classes of related problems.

The presentation reports 11.4 billion assets tracked, more than 29,000 miles of asset movements measured every hour, and an 88% time-to-insight reduction. The sensor-message count in the auto transcript is noisy, so it is not used as a primary scale claim.