Safety-critical
Allergies and contraindications.
Never dropped from context.A longevity platform whose AI knows when to refuse.
An AI-native longevity platform. A multi-model reasoning engine sits behind intent routing and safety guardrails, answers from the member’s own labs, documents and wearable data through a vector store, and carries a five-layer memory of what matters about them across every conversation.
Memory layers
5
Safety · clinical baseline · active plans · longitudinal trends · the human element
Backend modules
46
NestJS modules behind one authenticated API
ML service
138k
Lines of Python across 593 files
Member app
30.4k
Lines of React 19 + Vite
Every value, lab result and clinical detail in these screenshots is blurred. Sumero holds real member health data; only the structure of the interface is shown here.
01
The part most teams overpromise
The difference between a demo and something you can put in front of real users is not answer quality on a good day — it is behaviour on a bad one. The screen above is the bad one: contradictory source data, a question that cannot be safely answered, and a model that has to notice.
Every message passes through a control layer before generation — intent classification, out-of-scope and emergency detection, then safety guardrails.
The routing layer classifies what a message actually is (a question, a request for coaching, a data lookup, an emotional moment), detects emergencies before a response is generated, and adapts depth and tone to the situation. A fast model handles routing and classification; a heavier one handles clinical reasoning. Models run locally for cost control and privacy with cloud fallback for resilience, and providers are swapped by configuration — the client is never locked in.
02
Five layers of memory
Generic chat forgets. A longevity product cannot: the whole proposition is that it gets smarter about one person over years. Memory is split by what the fact is for, because the layers have different retention, different retrieval priority and very different consequences if wrong.
Clinical baseline
History, conditions, medications — the standing picture of the member.
Active plans
Current protocols and interventions the member is actually on.
Longitudinal trends
Biomarker movement over time — the direction, not just the latest value.
The human element
Goals, preferences and communication style — what makes the answer land.
Grounded, not remembered
Uploaded PDFs are processed, chunked, embedded into Qdrant and searched on demand, so answers cite the member’s real documents and lab results instead of model recall.
03
The screens
04
Engineering notes
Platform shape
- Two cooperating services. A NestJS API is the only thing clients talk to; it proxies AI calls into an internal FastAPI service over an internal-key + user-identity header contract. The ML service is never publicly reachable.
- Authenticated by default. The auth guard is registered globally, so every route is protected unless a handler explicitly opts out — the safe direction for a health product.
- Throttling is global too, with stricter per-endpoint limits opted into where abuse actually matters.
- Queues and infrastructure: Redis-backed BullMQ for background work, Qdrant for vectors, Postgres for records. The API refuses to boot if Redis is unreachable rather than degrading silently.
Posture and integrations
- HIPAA-grade posture: httpOnly cookies with cross-site-safe flags in production, security headers and HTTPS redirect at the edge, and an explicit CORS allow-list from environment.
- Wearables, labs, payments, video, email and cloud storage all integrate behind the same service layer, so the member’s picture is assembled from the tools they already use.
- A clinician is always one click away — the product routes to a human rather than pretending to be one.
- Deployment is a single scripted path: build, provision the Python environment, then run both processes side by side under a process manager.