Grounded, not generated
Every clinical answer cites the past visits it pulled from. No retrieval, no answer.
Research library
Show your work.
The AI inside MediSero+ is held to the same standard as a clinical study. Methodology, benchmarks, validation studies — published openly, with reproducible methods and honest about what's measured vs what's still in-prep.
Open methodology · Reproducible benchmarks · Citing sources by default
Design principles
Six rules. Every model decision goes through them.
When we add a feature, we run it past these. If it can't pass all six, we don't ship it.
Drafts only
Clio never auto-saves to a patient record. The doctor signs. Always.
Workspace-scoped
Each clinic's data is isolated. No cross-clinic learning without explicit consent + anonymisation.
Audit-first
Every Clio output is timestamped, attributed, and traceable to source. Forever.
Edit-in-place feedback
Doctor edits are the training signal. Local, per-clinic. Your edits make Clio yours.
Toggle anything
Master switch off. Per-screen off. Per-feature off. Privacy is a setting, not a contract clause.
What we've measured
Honest benchmarks. Reproducible.
Every number below is reproducible from the corresponding methodology post. No marketing-only stats. If a number isn't here, we haven't measured it yet.
9.7%
WER on code-switched dictation
Down from 31.6% baseline. Measured across 4,318 dictations from 22 pilot clinics.
Source · methodology post
89%
Top-3 differential accuracy
On 1,200 anonymised primary-care cases. Vanilla LLM scored 61% on the same set.
Source · methodology post
0.4%
Hallucination rate
With RAG + cross-encoder rerank. Vanilla LLM: 14% on the same benchmark.
Source · methodology post
1.3%
Drug-name error rate
Post-pipeline. Down from 18% baseline via clinical lexicon biasing.
Source · methodology post
Methodology — public posts
Two long-form posts. Full architecture, full numbers.
These two cover what's currently measured and shipped. Formal whitepaper versions will follow as the methodology stabilises through 2026.
How Clio cites past visits — RAG in clinical AI
Full architecture: pgvector inside the workspace, two-stage retrieval with cross-encoder rerank, prompt template, post-processing verification, citation rendering.
Read post
Speech-to-SOAP in Hinglish — measuring what works
WER per language across 4,318 dictations. Three failure modes (language ID flips, numerics, drug names) and the four-step pipeline that fixed them.
Read post
Papers & benchmarks
Six write-ups planned. Two public, four in prep.
Most methodology starts as a public blog post then matures into a formal whitepaper or peer-reviewed submission. We publish in the order we measure.
How Clio cites past visits — RAG architecture for clinical AI
pgvector inside the workspace. Chunked on every Visit, Rx, and lab upload. Retrieval scoped to the patient + similar cases, never cross-clinic.
Multi-language voice-to-SOAP — WER per scenario
Word-error rate measured on 4,318 dictations across English, Hindi, and code-switched Hinglish. Open methodology, reproducible from the corpus structure.
Hallucination & refusal rates — Clio vs frontier baselines on clinical Q&A
Expanding the rag-in-clinical-ai post into a formal write-up with extended baselines (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro), confusion matrices, and ablations.
Speaker diarisation in clinic recordings — DR vs PT vs family
Labelling up to four speakers per consult, surfacing ambiguity rather than guessing. Pilot data being collected through Q2 2026; write-up follows.
Right-to-erasure in a RAG-indexed clinical record
DPDP Act 17 / GDPR Article 17 right-to-erasure when a patient's embeddings are scattered across a workspace index. Cryptographic deletion via key-rotation.
Clinic-native model fine-tuning — three months in
Per-clinic fine-tunes vs zero-shot frontier models on note quality. Pilot underway with two specialist clinics. Target submission to a peer-reviewed journal.
Model stack
What's under the hood.
Transparency about what powers each Clio capability — provider, model, retraining cadence. We retrain or swap any of these freely; this table reflects the current production stack and updates whenever we ship a meaningful change.
| Capability | Provider · Model | Updated |
|---|---|---|
| Voice transcription (first pass) | Deepgram Nova-2 with confidence routing | Quarterly |
| Voice transcription (low-confidence segments) | Whisper-large fine-tuned + clinical lexicon biasing | Quarterly |
| Note structuring (SOAP / Brief / Detailed) | GPT-4o-mini · GPT-4o for hard cases | Monthly |
| Embedding for RAG retrieval | all-mpnet-base-v2 fine-tuned (EN + HI pairs) | Quarterly |
| Vector store | pgvector (per-workspace namespace) | Continuous |
| Reranker | Cross-encoder ms-marco-MiniLM-L6-v2 | Quarterly |
| Drug research (monograph + interaction) | Internal RAG over RxNorm + DrugBank · GPT-4o-mini | Weekly |
| Clinical Q&A (Clio chat) | RAG over workspace · GPT-4o-mini · GPT-4o for >1k tokens | Monthly |
| Suggested tasks | RAG + reasoning model · per-visit prompt | Monthly |
Sub-processor list (every third party that touches your data, with regions and contract dates) is on /compliance (coming soon).
Data policy — short version Do you train on customer data?
Is data sent to OpenAI used to train their models?
Where do you store embeddings?
Can I see my own retrieval logs?
What we do — and don't do — with your clinic's data.
Do you train on customer data?
No. Default. Per-workspace embeddings are used for retrieval only — never for cross-clinic model training. Cross-clinic learning requires explicit opt-in and is anonymised before any training set inclusion.
Is data sent to OpenAI used to train their models?
No. Our OpenAI Enterprise contract has a no-training-on-customer-data clause and zero log retention. Their API tier we use does not retain inputs or outputs beyond the request lifetime.
Where do you store embeddings?
In pgvector inside Postgres in your region (AWS Mumbai for India by default). Embeddings are deleted alongside their source records on right-to-erasure within 30 days.
Can I see my own retrieval logs?
Yes. Every retrieval that produces a Clio answer is logged to your workspace audit log: query, retrieved chunk IDs, ranking scores. Available in Settings · Compliance · Audit log.
Full policies at /privacy and /compliance (coming soon).
Reach out
Researchers, regulators, journalists.
If you want raw access to a benchmark, source data for a paper, or a deeper conversation about methodology — write to research@medisero.com.