Agentic AI for drug safety, with a human in every loop
Generative and agentic AI runs through all seven Nirnāśā services. It reads reports, codes events, detects signals, drafts reports and investigates deviations. Your experts review, override and sign every decision, and each step is logged.
From raw report to signed decision
The same five-stage pattern runs through every Nirnāśā service.
Understand
OCR for scans, speech-to-text for calls, and parsing for PDFs, Word and email.
Reason
Language models extract, classify, code, detect and draft, grounded in your data.
Verify
Confidence scores, model consensus and rule checks flag anything uncertain.
Human decides
Experts confirm, override with a reason, and e-sign the outcome.
Log & learn
Model, prompt version, confidence and overrides are kept for audit.
Seven services, one AI approach
PV Intake
- AI extraction with confidence scores
- Four-way classification with model consensus
- Literature and social-media AI triage
ICSR Suite
- Serious / non-serious triage
- AI-suggested MedDRA and WHO-Drug codes
- Narrative drafting and duplicate checks
Aggregate Manager
- AI drafting of PSUR, PBRER and DSUR sections
- Sections filled from product and case data
- Author edits always win
Signal Management
- AI triage copilot with citations
- Seven specialised agents
- Statistical detection: PRR, ROR, IC, EBGM
PSMF
- Six human-governed agents
- Change-impact and consistency checks
- Ask questions of your master data
QMS
- Agentic root-cause and CAPA drafting
- Similar-event detection and risk scoring
- AI audit-trail review
Voice Connector
- Real-time speech recognition
- Answers grounded in your documents
- Natural text-to-speech voices
See it on your data
We'll run the AI on anonymised sample cases or literature so you can judge the results yourself.
Book a demo
Specialised agents that work as a team
Instead of one general chatbot, each task has its own agent, with clear inputs, clear outputs and a person accountable for the result.
- PSMF: six agents for regulatory text, evidence, change impact, consistency, Annex I logbook and inspection preparation.
- Signal Management: seven agents for data intake, detection, WHO-UMC causality, lifecycle, reporting, demographics and normalisation.
- QMS: an investigation pipeline that parses, gathers context, checks data integrity, then drafts the RCA and CAPA.
- Orchestration: agents run one at a time, chained as a pipeline, or in parallel.
AI you can defend to an inspector
Regulated work needs AI that is explainable, controlled and recorded. These principles are built into every service, not added afterwards.
Human in the loop
AI proposes. Qualified people confirm, override with a recorded reason, and e-sign.
Explainable outputs
Confidence scores, sources and plain-language rationale for AI suggestions.
AI decision log
Model, prompt version, inputs, outputs and overrides are kept for audit.
PHI masking
Names, dates of birth, phone numbers and other identifiers are removed before text reaches a language model.
Safe fallback
If a model is unavailable, machine-learning and rule-based methods take over, and their results are flagged.
Grounded in your data
Answers come from your documents and records. Assistants that report figures read them from the data, never guess.
AI technology we build with
Questions about our AI
No. AI extracts, classifies, codes, drafts and flags. A qualified person confirms or overrides every suggestion, and approvals are e-signed.
Our services work with OpenAI models such as GPT-4o, Anthropic Claude and AWS Bedrock models. Providers are configurable, and some modules support an OpenAI-compatible or self-hosted model.
Personal identifiers such as names, dates of birth and phone numbers are removed before text is sent to a language model. We agree hosting regions and model providers with your security and privacy teams.
Work keeps moving. Machine-learning and rule-based methods take over, and their results are flagged so a person confirms them.
Yes. Suggestions come with confidence scores, their source and, where available, a plain-language rationale. The model and prompt version are logged for every decision.
Put human-governed AI to work on your safety data
See extraction, coding, signal triage and report drafting running on anonymised samples of your own data.