Here's what I'm proposing for our crime data format:
{
“_id”: String (autogenerated),
“AgencyID”: String,
“AgencyName”: String,
“CaseNumber”: String,
“CriminalOffense”: String,
“DateReported”: DateTime (YYYY-MM-DDTHH:MM:SS),
“Description”: String,
“Location”: String,
“OccurenceDate”: DateTime (YYYY-MM-DDTHH:MM:SS),
“Disposition”: String,
“Lat”: float,
“Lng”: float
}

• _id is autogenerated by CouchDB if you need, or could use a unique id like CrimeCodeID
• Coordinates will need to be converted or geocoded to lat/lng
• Location should have the city/state added to the street address
• CriminalOffense (CrimeCode) should probably be standardized to some extent

Does anybody have any thoughts on whether we should essentially keep all of the original data we scrape and then add the standardized fields separately?

It seems that we've taken on the challenge of finding the safest path, which is a pretty difficult problem. As an alternative, we could get the path from Google and just give it a safety score. Here's a demo where a marker is animated through a path. We could take a number of positions along the route, like that demo seems to do, and score them each. Then we could aggregate the scores and present a final score. We could then set thresholds and make recommendations (it's ok to walk, you should probably get a ride, etc).

If we do this, we just need to make a number of APIs that take a point and return a yes / no answer e.g. was there a recent crime here, is there a sidewalk here, etc.

Future iterations could show the dangerous portions of the walk on a map and allow the user to manually change the route and re-score.