MondoTwin Tech · The Twin Stack
Know. Predict. Work.
Act. Learn.
A company is what it knows, how it judges and how it acts, and it lasts by learning from what happened. The Twin builds that cycle as one line of software, for one field at a time.
Five layers
What runs
in each layer.
01 · Know
Twin Ontology
A reader with limited time answers better from a compiled world than from raw sources, as long as every fact opens down to its origin.
Give company data, documents and entities meaning. Connect the relationships and preserve their sources.
What runs here
- Capture. Every source read with care; every record kept with its origin and version.
- Identity. One drug, trial or disease under many names becomes one entity. Merges reversible, splits recorded.
- Meaning. Mentions linked to vocabularies; every extracted value tied to the passage it came from.
- Facts. Claims from every source kept side by side, conflicts visible, dated twice: when true, when known.
- Releases. Immutable, promoted as a whole, cleared before publication.
Pharma today
Public trial registries in the US, Europe, the UK, Australia, Japan and at the WHO; ICD-11, Orphanet and Mondo; DrugBank, GSRS and UNII; approvals, guidance, the literature.
02 · Predict
Twin Models
Once the record is clean, the regularities of the field can be learned from it. A prediction carries the same provenance as a fact.
What runs here
- Specialised models trained over the ontology. Today: the chance a trial succeeds, one calibrated model per phase, tested on later years than it learned from. Next: how long a trial takes.
- A decision cascade: rules, then a trained scorer, then a small model, then a large one. An answer is accepted only by a named rule, calibrated on gold.
- A model gateway: every call recorded and replayable, under spend caps. Models change; knowledge and calibration stay.
Pharma today
Trial success in Phase 2 and Phase 3; synergy and dose-response for combinations.
03 · Work
Twin Agents
An agent’s work is only as good as the world it reads and the tests it passes.
Configure task-specific agents with the right context, tools and permissions. Evaluate their work before expanding autonomy.
What runs here
- One agent framework: structured agents, one model resolver, every run traced.
- A skills engine: a procedure written for a coding agent runs unchanged on the platform, every tool call logged.
- Evaluation: gold sets, a judge, and named human review before work reaches a client.
Pharma today
Asset analysis, regulatory drafting, evidence agents, and skills for candidate assessment and combination design.
04 · Act
Twin Works
A department is its procedures and its rules. Written down, they can be run by an agent and governed by the people who answer for them.
Connect memory, workflows and accountable decisions across the company. Integrate with agreed existing systems.
What runs here
- Departments as repositories: procedures as skills, rules by risk. Routine acts delegated; sensitive acts proposed; money and legal acts performed by a person.
- Truth read live from the system that owns it. Nothing cached.
- A runtime: a message on WhatsApp, Discord, Jira or the web starts a session in that department, with only its access. Every tool call audited.
- The platform: organisations under agreement, workspaces, rooms where a client reviews an asset; the Twin reachable from Claude, ChatGPT, an API and chat.
Pharma today
Business Development first: the analyst cockpit, the asset registry, the clinical-studies browser.
05 · Learn
The Loop
When an agent fails, repair the world it read, not the agent.
What runs here
- Every change is replayed against recorded answers and judged on sealed gold sets: accept, reject or cannot tell.
- Protected metrics the change cannot see. A rejected change leaves a record.
- Next: agents’ tool calls become traces; recurring failures are followed through lineage to the decisions behind them, and a repair is proposed, replayed and gated.
Pharma today
Gold sets for trial identity, drug identity, disease coding and trial outcomes.
Across the stack · Trust
Provenance to the passage. Access decided per source and cleared before publication. Named human review. Agents propose, people dispose. Every act logged.
Across the stack · Cloud
AWS in Frankfurt. Infrastructure as code, deploys keyed by content, architecture boundaries tested.
One line of software.
Many fields.
A field is a pack: sources and their terms, vocabularies, entity types, predictors, agents and skills, a first department. Pharma is the first. The stack holds no pharma word and was reviewed against other fields, from energy to law, before pharma’s pack was written.
The difference
A data platform is sold for a customer’s team to build on. MondoTwin builds and runs the Twin, department by department. The department layer follows an open specification.
A focused first function.
Choose the starting point of your project.