The platform under every Roshan AI product.
Roshan AI is a platform company. Every product we ship (ShifaMind today, more in development) reads from the same ingestion layer, runs on the same clinical encoders, and explains itself through the same concept reasoning layer. This page is the stack.
Ingestion
Clinical data, normalized into something models can read.
Discharge summaries, progress notes, ED notes, structured EHR fields, imaging reports. Notes arrive over API or batch; the ingestion layer normalizes formatting, strips or preserves PHI per customer policy, and produces the canonical text representation downstream models consume.
- Free-text notes and discharge summaries
- Structured EHR exports (FHIR-flavored)
- PHI tagging and de-identification (configurable)
- Per-tenant data isolation
Models
Clinical encoders trained on real clinical text.
BioClinical ModernBERT-base today, with larger backbones under evaluation. The encoder is paired with a learnable concept-query bank trained against an explicit set of clinical concepts: the same concepts a coder or clinician would name out loud. The encoder is the substrate; the concept-grounded representation is the contract.
- BioClinical ModernBERT-base, 8,192-token context
- 160 grounded clinical concepts
- Multiplicative Concept Bottleneck (MCB)
- Larger backbones under evaluation
Reasoning
A concept bottleneck every prediction must flow through.
The reasoning layer takes an encoded note, activates the concepts present, and routes prediction signal exclusively through those concepts. No prediction is produced without the concept evidence that supports it. The bottleneck is the feature: it forces explainability by construction.
- Concept activation per note
- Code prediction grounded in concepts
- Confidence and alternatives
- Concept-Supported True Positive Rate (CSTPR) telemetry
APIs
One integration shape, many products downstream.
Every product on the platform exposes the same response shape: predictions, evidence, concepts, alternatives. Integrators write to a single integration surface; products plug in mechanically. The API is the boundary between Roshan AI infrastructure and the workflows clinicians actually use.
- REST endpoints, JSON responses
- Token-based auth
- Per-prediction audit logs
- Webhooks for async pipelines (roadmap)
App surface
The first product on the stack, today.
ShifaMind is the first consumer of the platform: a coder and reasoning workspace clinicians use directly. The same platform layers power the products that come next. Adding a new product is a content and policy exercise, not a re-architecture.
Integrate ShifaMind today. Build on what's next.
The same platform that powers ShifaMind will power the products that come after it. If you're building a clinical workflow that needs grounded reasoning, talk to us early. Partner deployments shape the roadmap.