public records about Boston homes, remembered for the people who live in them.
HomeSafe pairs the city’s own data with a renter’s private, dated journal and an agent that answers questions with receipts: every claim cites a source it actually read, and anything it can’t prove gets deleted before you see it.
Look up any Boston address, right now
Violations, permits, and housing complaints, each labeled with how confidently it was matched, and what it does not prove. A permit is never presented as a repair.
How it helps a renter hold a landlord accountable
A journal that stands up
Dated, private notes. Photos are described by you, never analyzed by the AI, and location data is stripped before an image leaves your phone.
Your building’s paper trail
One real six-unit building in Roxbury carries 53 heat complaints and one permit. One complaint is a story; fifty-three is a pattern an inspector can act on.
Answers with receipts
The agent reads your notes, the city’s records, and the Massachusetts sanitary code, and shows exactly what it read. If it cannot cite a claim, it deletes the claim and tells you it did.
The hackathon criteria, exceeded
| Criterion | Required | What HomeSafe ships |
|---|---|---|
| CockroachDB tools | at least 2 | All 4: distributed vector indexing (1024-dim, consent-filtered), managed MCP at build time with findings filed as feedback, ccloud for the two-login security model, and the Agent Skills repo doing load-bearing privilege work. |
| AWS services | at least 1 | 3: Bedrock (agent + embeddings via a role that can invoke exactly two models, no keys in the build), Amplify Hosting, S3. |
| Store, retrieve, act on memory | the challenge | All three verbs, live: the agent writes its own conclusions back to memory, recalls them next session by meaning, and drafts approval-gated tasks. Six kinds of memory in one cluster: records, journal, vectors, agent diary, task state, audit log. |
| Memory not an afterthought | the judging bar | Memory that proves itself: every answer ships a receipt of what was read, written by the retrieval layer. The model cannot author it. The receipt is also the audit row and the validator’s source of truth. |
| Production readiness | secure, resilient | Privacy by missing GRANT, prompt injection tested live (nothing leaked; the withheld count moved), notes survive model outages, revoked memory erased by the database’s own row-level TTL, follower reads on the million-row timeline. |
| Real-world impact | meaningful use case | Boston renters vs. undocumented housing conditions, with the entire Massachusetts sanitary code retrievable by meaning and cited to the section, plus referrals to legal-aid guides. |
191
tests passing, several against the live cluster and real model calls
138–155 ms
consent-filtered vector search over seeded memories
87
Massachusetts housing rules in memory, each linked to its source
0
private items leaked under live prompt-injection attack
See the receipt for yourself
Sign in with just a name, open a case, ask a question. Then open “Why do I remember this?” and read exactly what the agent read.
Open the live demo