Loading...
Each agent's output bounds the valid output space of every downstream agent. The grade that exits the engine is the only grade consistent with every prior agent's findings. Citation-bound by design. Real-time across every site. Two patents filed on the architecture.
837+ CTCAE criteria across v5.0 and v6.0. MedDRA v28.0 LLT codes attached to every graded event. 42 oncology regimen profiles. WHO-UMC and Kramer attribution with visible reasoning. 21 CFR Part 11 e-signature trail. Human review on every signed record.
The cascade is the architecture. The cascade is also the moat.
In the pipeline. Each output bounds every downstream agent's input space. The cascade is the architecture.
Filed. The pattern is the moat, not the implementation. The cascade cannot be replicated by training a better classifier or wrapping a better prompt.
Without citation. Architectural constraint, not runtime filter. The engine cannot return a grade without a source sentence and a matching CTCAE criterion.
Most clinical AI systems are classifiers. A clinical note enters; a label exits. The Burna engine is not a classifier.
It is a pipeline of twelve specialized agents, each performing a specialized reasoning task and producing a structured output. Each agent's output narrows the valid output space of every downstream agent. By the time the engine reaches the grading decision, the output space has been bounded by eleven prior decisions. The grade that emerges is the only grade consistent with every prior agent's findings.
The cascade is the architecture. The cascade is also the moat.
The architectural thesis becomes a property when twelve agents coordinate. Each agent's output narrows the valid output space of every downstream agent.
Signal Acquisition opens with roughly fifty candidate phrases drawn from the clinical note, the laboratory results, the medication list, and the regimen protocol. Assertion Verification reduces to about thirty positive assertions. Entity Resolution to about twenty. Pharmacological Reasoning to under fifteen entities carrying a pharmacological substrate in the patient's record. Differential Analysis to about ten events the engine can defend as drug-related. By the time Causal Attribution runs WHO-UMC and Kramer against the evidence record, fewer than five candidate attributions remain. The grade that exits the engine is the one consistent with every prior agent's findings. Not a probability distribution. A single grade.
Representative magnitudes for a typical Phase 2 oncology adverse event workup. Actual counts vary with note complexity, regimen, and prior cycle history.
Signal Acquisition, Assertion Verification, Entity Resolution, Pharmacological Reasoning, Taxonomic Intelligence, Differential Analysis, Temporal Awareness, Documentation Integrity, Evidence Synthesis, Causal Attribution, Confidence Calibration, and Regulatory Encoding. Each agent operates under the constraints of every upstream agent.
Reads the clinical note, the laboratory results, the medication list, the prior cycles, and the regimen protocol. Identifies candidate phrases that may indicate adverse events.
Validates which candidate phrases are positive assertions of clinical fact. Distinguishes assertions from negations, hypotheticals, and family history references.
Disambiguates clinical entities against the patient's longitudinal record, not against a generic medical knowledge graph.
Cross-references the working set against active regimen, prior regimens, comorbid medications, and the 42 oncology regimen profiles.
Maps entities to the appropriate CTCAE criterion across 837+ terms in v5.0 and v6.0. An entity that does not have a defensible criterion does not advance.
Distinguishes drug-related toxicity from comorbidity-related findings, disease-related findings, and incidental findings.
Tracks exposure, onset, peak, resolution, dechallenge, and rechallenge. Builds the temporal evidence chain attribution requires.
Detects copy-forward documentation, stale references, and contradictions across the longitudinal record. Prevents an event documented incorrectly from compounding into a Grade 3 finding.
Assembles the clinical evidence supporting each candidate adverse event into the evidence record that travels with every grading decision.
Runs WHO-UMC and Kramer causality algorithms against the evidence record. Per-drug probability scores with the algorithmic reasoning path visible.
Assesses the strength of the evidence chain. Flags low-confidence cases for clinician review. A feature the clinician sees, not a number the engine optimizes against.
Encodes the final decision into the formats downstream systems require. MedDRA v28.0 LLT code for pharmacovigilance and regulatory submission (see /meddra-coding for the coding capability in full). CTCAE grade for the trial database. E2B(R3) for ICSR. The clinician's signature is captured on the encoded record, not on the raw output.
Properties of the retrieval architecture. Not transparency layers added after the fact. Not dashboard preferences. Structural.
Citation
Every grade the engine produces requires a source sentence in the clinical note and a matching CTCAE criterion in the retrieval index. Without both, the engine returns “no grade available.” The constraint is the architecture.
Constraint
The engine cannot produce a grade outside the defined CTCAE set, cannot skip citation, cannot contradict its own upstream findings. These are properties of the architecture, not constraints applied at output.
Real-time
Every grading decision propagates instantly across all stakeholders. Grade 3+ events trigger immediate alerts. In multi-site trials, grading inconsistencies surface in real time, not at the next monitoring visit. Built from day one. The engine's real-time capability is a property of the substrate.
Engineering specification translated for sophisticated readers. The page demonstrates the architecture; it does not pitch the architecture.
Built on a twelve-agent cascading constraint pipeline. Two patents filed on the pattern.
Three pre-negotiated document sets. The 5-page Architecture and Implementation Brief covers all three. Available at security@burna.ai.
Download the 5-page brief →Multi-tenant SaaS under BAA, customer-isolated single-tenant inside your VPC, or sponsor-tenant Protocol Safe inside your AWS, Azure, or GCP. The engine is identical across all three. The audit trail is identical. Two patents filed on the pipeline.
Regional cloud at launch: US, EU (Frankfurt, Dublin, Paris GDPR-resident), UK, Canada, Australia, Japan, Korea, Singapore, UAE, KSA. Data does not cross jurisdictions without explicit configuration. The grading model only sees de-identified clinical text. No model training on customer data without explicit written consent. Ever.
MSA, QA, DPA, subprocessor list, insurance certificates, right-to-audit clause, source code escrow option, and 90-day exit and data export provisions are pre-negotiated, single-document downloads. 21 CFR Part 11 validation package ships with every engagement. GVP Module VI alignment documented. EDSTP eligible.
“The constraint is the architecture. The engine structurally cannot produce a grade outside the defined CTCAE set, cannot skip citation, cannot contradict its own upstream findings. These are not features. They are properties.”
From the Architecture and Implementation BriefBurna AI, technical document, May 2026 revision
The engine is not a single model that someone can swap for a competitor's model. It is twelve agents, cascading constraints, citation-forced output, and a patented architecture. The four objections that recur in evaluation conversations have specific architectural answers.
A single model does classification. A clinical note enters; a label exits. The Burna engine does reasoning under constraint. Twelve agents. Each output bounds every downstream agent's input space. The grading decision at the end of the chain is the only decision consistent with every prior agent's findings. That is not classification. That is architecture.
An LLM wrapper is a single generative model with a prompt and a retrieval step. The Burna engine structurally cannot produce a grade outside the defined CTCAE set, cannot skip citation, cannot contradict its own upstream findings. The constraint is enforced at the architecture level, not at the prompt level. A wrapper hallucinates because nothing in its architecture prevents it from hallucinating. The Burna engine cannot hallucinate because the architecture does not permit it.
The pipeline spans twelve agent categories. Each agent is independently developed, tested, and constrained. The cascade between agents is the patented contribution; two patents are filed on the architecture. A weekend project produces a single classifier or a single LLM call. It does not produce a constraint pipeline.
Veeva Vault Safety, Argus, ArisGlobal, and Medidata Detect operate downstream of the attribution decision. They manage cases after the investigator has already decided what caused the adverse event. The Burna engine operates upstream, where the investigator is making the attribution decision. Nobody else helps the investigator make that decision at the point of clinical encounter. The white space is structural, not feature-level.
The patented contribution is the cascading constraint pipeline itself: the pattern in which twelve specialized agents coordinate to produce a grading decision that is bounded by every prior agent's findings. The patents cover the architectural pattern. Implementation details remain proprietary.
The architecture is the moat. The engine cannot be replicated by training a better classifier, fine-tuning a better LLM, or wrapping a better prompt. The moat is the cascade.
The engine produces a suggested grade with its citation, its evidence chain, its named-algorithm attribution, and its confidence calibration. The clinician reviews. The clinician modifies, approves, or rejects.
Only the approved, signed version becomes the record. Every modification is logged. Every rejection feeds back into site-specific tuning under explicit governance review. The engine's job is to produce a defensible suggestion. The clinician's job is to make the decision. The architecture is built around the clinician's authority, not around the model's autonomy.
The 5-page Architecture and Implementation Brief covers the engine architecture, the deployment modes, the data flow diagram, the validation roadmap, and the procurement documentation. Available at security@burna.ai.
The 2026 design partner cohort accepts pilot scoping conversations through the end of Q3 2026. Pilots that begin in Q4 2026 ship co-authored manuscripts in late 2027.
For investors, the engine page is the technical companion to the investor brief. For pharma and CRO conversations, the Protocol Safe brief extends the engine architecture to sponsor-tenant deployments.
Subscribe to the quarterly safety and data quality briefing. Validation readouts, platform updates, and conference timing only.
No tracking beacons, no marketing email cadence.