Independent public project
How a corrected crash report reaches the California map.
Meru built and operates Evidence for Zero, a public map of reported California crashes. Behind each point is a record that can change after it first appears.
One collision can appear in more than one release.
When California revises a collision report, a fresh row in a download can look like a second crash. Meru keeps the source versions, matches the revision to the original collision ID, and updates the affected map data. A researcher comparing neighborhoods can count that collision once and trace the change back to its source.
Explore the result on the live Evidence for Zero map, or follow one fictional record through the same decision in the example below.
- Business problem
- Public crash data was fragmented across records, years, revisions, and geography.
- System built
- A cloud data pipeline, prepared analytics warehouse, API, and interactive map.
- Decision value
- A clearer view of where reported activity is changing and where review should begin.
Follow one record from source to local view.
Follow a fictional collision from its original source row to an amendment, typed fields, and a prepared local view. This schematic example is separate from the live public map.
About this example
Fictional data. Prebuilt outputs, not live AI. No client data is uploaded and no messages or changes are sent to a client system. Interactions stay in this page until reset or reload; where offered, downloads save only the authored sample files. The interface illustrates the working patterns described on this page.
Record lineage
Meru · implementation workspaceTaylor Quinn · data reviewer
Update the record without counting it twice
Step 1 of 3: Match the event
- Match the event
- Review the amendment
- See the updated view
A new export includes a collision already in the file. Taylor needs to recognize the amendment before it reaches the local brief.
Original export · October 12
DEMO-COLLISION-001
Incident: October 10 at 08:20
Later export · October 16
DEMO-COLLISION-001
Incident: October 10 at 08:20
The stable ID links both rows to one collision. The later version supplies new detail.
Inspect the original and amended rows
EXAMPLE-EXPORT-1012 · row 1 · Published 2026-10-12
Collision key: DEMO-COLLISION-001. Incident: 2026-10-10 08:20. Locality: Example Borough.
Road fields: EXAMPLE AV / Sample Lane. Severity code: UNSPECIFIED. Injury count: not supplied.
Schematic grid position: not supplied. Speed: not supplied.
These are authored sample values, not a California collision or real geographic coordinates.
EXAMPLE-EXPORT-1016 · row 1 · Published 2026-10-16
Collision key: DEMO-COLLISION-001. Incident: 2026-10-10 08:20. Locality: Example Borough.
Road fields: Example Avenue / Sample Lane. Severity code: MINOR. Injury count: 1.
Schematic grid position: 64, 42. Speed: not supplied.
These are authored sample values, not a California collision or real geographic coordinates.
A changed source row should change the record, not inflate the count.
- 01Retained source
Two versions
Published October 12 and October 16 for the same collision key. - 02Prepared data
One collision
A reviewed revision replaces the prepared record, not the source history. - 03Local view
No map point yet
Missing position stays missing. No real geography is represented.
Next actions
1 ready to reuse2 to review1 still to obtain
Identify the same event in both downloads
BeforeCompare separate exports to decide whether the later row is new.
Both source rows retain DEMO-COLLISION-001 and the October 10 incident time.
Original and amended sample exportsIn the recordReview the amendment
BeforeCopy changed values into a working sheet and keep track of what changed.
A field-level comparison separates the received amendment from the still-active original.
October 12 original / October 16 revisionReview neededDecide whether the record can be plotted
BeforeRecheck the location fields before adding a point to a map.
The amendment supplies an explicitly fictional grid position. It enters the schematic only after review.
Amended row · schematic coordinate fieldsReview neededKeep unsupported conclusions out
BeforeDistinguish empty source fields from zero or an analyst's assumption.
Speed remains unknown. Neither the source nor this example establishes fault, safety, or case value.
Source fields and normalization notesFollow-up needed
These are work items in this fictional file, not a count of calls or hours saved. Reuse dated evidence; contact people for changes, missing facts, and decisions that still need a conversation.
Explore reported crashes in Los Angeles.
This is the live project, separate from the fictional example above. The embedded view is filtered to January through May 2026 in Los Angeles. Open the full product to change the date, geography, and available collision filters.
How the public data became usable.
Cleaning was treated as a versioned system, not a one-time spreadsheet task. Every stage keeps a path back to the source.
- 01
Preserve the public record
California CCRS crash, party, and injury records are ingested from data.ca.gov. Normalized source columns and source metadata remain available for audit and backfill.
- 02
Reconcile revisions
Stable collision IDs, normalized row keys, hashes, and source watermarks identify changed records. A rolling overlap catches late edits instead of treating each download as final.
- 03
Create typed facts
Dates, coordinates, severity, parties, injuries, vehicles, fault indicators, and location fields are cast into separate collision, party, and injury facts with explicit null handling.
- 04
Standardize inconsistent labels
Vehicle makes, severity codes, city labels, road names, and commercial-vehicle rules are mapped through versioned reference tables rather than overwritten in place.
- 05
Resolve geography
Source coordinates are matched to California ZCTA boundaries. Normalized road pairs can be checked against OpenStreetMap references when location context needs careful reconstruction.
- 06
Serve prepared signals
Partitioned BigQuery marts prepare map points, ZIP-by-day summaries, vehicle signals, collision facets, and hotspot shifts so the interface reads narrow tables instead of rebuilding joins on every request.
The product makes local demand easier to investigate.
The map does not manufacture intent. It gives teams a more disciplined place to start territory, content, outreach, and market conversations.
Plaintiff law firms
Where is case-relevant activity emerging?
Review recent crash patterns by geography, severity, vehicle context, and reporting narrative to inform market planning and investigative follow-up.
Media and growth teams
Where should local campaigns become more specific?
Compare ZIP-level activity and changes over time before allocating content, creative, outreach, or paid-media attention.
Executive teams
What changed in the market, and is it material?
Replace anecdotal market stories with a repeatable view of recent activity, historical baselines, reporting lag, and data completeness.
Analysts and operators
Can the source be trusted enough to act on?
Trace a signal back to normalized facts, source fields, geography methods, and documented limitations before it enters a decision.
Built to keep analysis out of the request path.
- Source
- California CCRS via data.ca.gov, DMV fleet snapshots, ZCTA boundaries, and OpenStreetMap road references
- Ingestion
- Scheduled delta sync with overlap windows, source watermarks, row hashes, and affected-collision reconciliation
- Warehouse
- Append-only raw records, typed BigQuery facts, versioned references, and partitioned serving marts
- Application
- FastAPI on Cloud Run with a React interface on Cloudflare
- Review model
- Signals support market and investigative review. They are not legal conclusions, safety ratings, or case-value predictions
Public data can become a proprietary operating advantage.
The defensible work is not the map alone. It is the process that keeps changing records traceable, makes geography queryable, and turns expensive joins into prepared answers. The same pattern applies to fragmented public, vendor, CRM, call, document, and operational datasets.
Important limitations
Records reflect reported collisions and can be revised or delayed. Missing coordinates and source-label variation affect some geographic views. Registration data describes vehicle stock, not miles traveled or risk. Review signals are not findings of fault, legal conclusions, safety ratings, or predictions of case value.
Questions about the data product.
What is The Crash Data Project?
The Crash Data Project is Meru AI’s independent civic research project. Its public product, Evidence for Zero, turns California collision, party, and injury records into maps and local profiles.
Where does the crash data come from?
The core records come from California's CCRS resources published through data.ca.gov. The system also uses California ZCTA boundaries, OpenStreetMap road references, and annual California DMV fleet snapshots for limited contextual comparisons.
How is the crash data cleaned?
The pipeline preserves raw source fields, reconciles revisions by collision ID and row hash, casts typed facts, standardizes labels through versioned rules, resolves geography, and validates freshness before preparing query-specific BigQuery marts.
How can crash intelligence support sales and marketing?
It can help professional services teams identify changing local demand, prioritize geographic research, plan market-specific content, and give executives a repeatable evidence layer for territory and campaign decisions.
Does the platform predict legal outcomes or crash risk?
No. The platform organizes reported public records for review. It does not determine liability, predict case value, measure vehicle safety, or substitute for legal, statistical, or investigative judgment.
Sources and author
Core collision, party, and injury records originate from California CCRS resources published through data.ca.gov. Geography and road context use public reference datasets. Meru preserves source metadata and documents transformations so prepared signals can be reviewed against the underlying record.
Suggested citation: Okabe, Jose. "How a corrected crash report reaches the California map." Meru AI, updated November 6, 2026. https://meruai.co/case-study/california-crash-intelligence
Author and review note
Written and reviewed by Jose Okabe, AI implementation strategist and enterprise systems architect. Jose designed and built the data pipeline, cloud warehouse, API, analytical methods, and interactive product described in this case study.
Last updated: November 6, 2026
Your scattered data may already contain the market signal.
Meru designs the pipeline, decision model, and working product that make difficult data useful to operators.
The Crash Data Project is an independent Meru AI civic research project that makes California public crash records easier to understand for residents, journalists, analysts, and operators.
Related: AI system integration