Trust & Security

1 · What this system is

The Location Copilot is a location-intelligence assistant for Thailand. Its flagship use case is EV charging-station site selection; the same engine answers site questions for other business types. Every answer is produced from two live data APIs plus one AI model, all called at question time:

2 · How we prevent AI hallucination

The model is not allowed to answer from its own memory. Controls are enforced in code, not only in prompts:

• Tool-grounded numbers. Every number, coordinate and place name must come from an API call made in the same turn. Ranking, scoring and distances are computed by our deterministic code — the model reports them, it does not invent them.
• The “Behind this answer” panel. Every answer ships with the full list of API calls that produced it (system, endpoint, result). Users can audit any answer themselves, on screen.
• “Not measured” is never rendered as zero. Supply data is always in one of three states — measured (real distances from a live search), empty (searched, none found), or unknown (not measured — the score falls back to demand-only and no distance is ever claimed).
• Mandatory geocoding. The model is forbidden from recalling coordinates; addresses and place names must go through HERE Geocoding, and unresolvable places are reported as not found — never guessed.
• Server-side output scrubbing. Responses are filtered server-side before display (internal identifiers are stripped) — a code-level guard, independent of the model's behaviour.
• Scope refusal. General-knowledge questions (news, history, coding…) are declined. If it cannot be traced to an API call, the system does not answer it.

3 · No internet access at answer time

When answering, the system can reach exactly two data sources: the HERE API and the DataProteins API. The model has no web-browsing or search capability in this application. Nothing is pulled from the open internet, so answers cannot be contaminated by unverified web content.

4 · Is customer data used to train the AI?

No — not by default, per Anthropic's published policy. This system calls the Claude model through the commercial Anthropic API. Anthropic's data policy states:

The stated exception is when a customer explicitly submits feedback/bug reports to Anthropic — this application has no such feedback channel, so no conversation is ever shared for training through it. If a user pastes business data (e.g. real sales figures) into the chat, it is used in-context only to answer that conversation: it is not written into DataProteins' analytical database and not used to train any model.

5 · What we store, and for how long

Accounts and audit entries can be deactivated/deleted by the administrator on request.

6 · Access control & application security

• Per-user accounts with roles (admin / user); accounts are created by the administrator only.
• Passwords hashed with PBKDF2-HMAC-SHA256 (200,000 iterations, per-user salt); never stored in plain text.
• Sessions via HttpOnly, SameSite cookies; only a SHA-256 hash of the session token is stored server-side; sessions are revoked on password change or account deactivation.
• Login rate limiting per source IP; failed attempts are logged.
• Server-held API keys (HERE / Anthropic / DataProteins) are never exposed to the browser — map tiles are proxied server-side specifically so the HERE key cannot leak.
• All traffic over HTTPS/TLS in production; per-user daily usage caps limit abuse and cost exposure.

7 · Scope & known limitations (POC honesty)

8 · Contact

Security questions and disclosure: dataproteins@gmail.com. We answer security review questionnaires as part of any evaluation.