A connected, permission-aware work graph
Profiles, competencies, teams, project structures, task activity, collaboration signals, commercial context and outcomes remain connected to their source objects and access boundaries.
Collty technology
Collty Intelligence Core
The private orchestration and learning layer that connects Collty data, domain algorithms, replaceable AI models and governed product decisions. It is designed to improve with real work while keeping users, evidence and human control at the center.
People & skillsTeams & relationsProjects & tasksSignals & outcomes
Context boundaryDomain algorithmsEvaluation & routing
Primary modelAnthropic shadow modeVersion comparison
Evidence-linked outputProvenanceHuman approval
Controlled learning loopFeedback and outcomes return through private evals, holdout testing and approval.
Architecture
A system around the model, not a wrapper around an API.
The model contributes interpretation and synthesis. Collty keeps the evidence model, domain logic, evaluation, permissions, provenance and activation controls inside its own platform boundary.
Algorithms structure the problem before an LLM sees it
Retrieval, ranking, capability coverage, availability, compatibility, workload, delivery and risk signals reduce ambiguity and prepare grounded evidence for the task.
Only task-relevant context crosses the model boundary
Context is minimized, structured and filtered. Sensitive fields, secrets and unrelated workspace data are excluded or redacted before a provider request is made.
External models are components, not the product brain
Provider adapters let Collty select a suitable model and compare alternatives in shadow mode without transferring product control or proprietary ranking logic to one vendor.
Outputs stay traceable to evidence and versions
High-value runs preserve provenance: the task, evidence boundary, algorithm and grader versions, provider, result status and the human decision that followed.
Outcomes improve the system through a gated loop
Feedback and project outcomes become evaluation signals. Candidate calibration is tested on holdout data, observed in shadow mode and activated only after human approval.
Learning and improvement
Every improvement has to earn its way into production.
A rating, accepted recommendation or completed project does not directly rewrite the ranking engine. Signals are aggregated into a controlled calibration process, tested away from production and reviewed before activation.
No silent self-modification.Production weights change only after holdout evidence and explicit human approval.
- 01Observe
Capture useful human choices, accepted actions and delivery outcomes as scoped learning signals.
- 02Evaluate
Run versioned private datasets through task-specific graders for assembly, planning and recommendations.
- 03Calibrate
Fit a candidate ranking version under bounded constraints instead of rewriting production behavior directly.
- 04Hold out
Compare the candidate with the current baseline on evidence it did not use for calibration.
- 05Shadow
Measure quality, drift, latency and cost without changing what users receive.
- 06Approve
A person reviews the evidence before a new ranking version can become active.
Data and privacy
Useful context in. Unnecessary data out.
Collty uses data to provide the requested workspace feature, ground a decision and improve internal quality through controlled signals. The external model is not the system of record and does not receive unrestricted access to the Collty work graph.
What can be used
- Authorized profile and capability evidence relevant to the task.
- Team, project, task and delivery signals inside the permitted workspace scope.
- Human feedback and project outcomes selected for controlled evaluation.
- Operational telemetry needed to measure quality, latency, errors and cost.
What stays protected
- Credentials, encryption material, payment details and secrets are excluded.
- Unrelated people, projects, chats and workspace records are outside the context boundary.
- Sensitive values are filtered or redacted before a provider request.
- Customer workspace content is not sent to model providers to train shared foundation models.
- Private eval cases, prompts, weights, thresholds and feature schemas are not public.
How decisions stay accountable
- Runs are associated with task, evidence and algorithm versions.
- Operational provenance favors identifiers and control metadata over raw content.
- Protected traces are encrypted when deeper investigation is necessary.
- Consequential changes preserve the human choice that approved or rejected them.
This page explains the AI-specific design boundary. Collty's broader handling of personal information is described in the Privacy Policy.
Production controls
Measured as a live system.
Quality cannot be separated from reliability and cost. Collty tracks the behavior of the complete decision path, not only the text returned by a model.
Quality and drift
Domain scores, critical failures and version comparisons reveal regressions before they become the new baseline.
Reliability and latency
Runs are monitored for duration, provider errors, retries and operational status across product workflows.
Durable execution
A durable outbox protects provenance and learning events, while distributed deduplication prevents equivalent AI work from being charged or executed twice.
Human control
New ranking weights and consequential project actions remain approval-controlled rather than silently changing production behavior.
QualityLatencyErrorsToken usageCostDrift
Standards and guidance
Designed in alignment with current responsible AI practice.
These frameworks inform Collty's engineering and governance approach. Alignment describes design intent and operating controls; it is not a claim that Collty is certified by NIST, ISO or a model provider.
NIST AI RMFGovern, Map, Measure and Manage
Collty applies lifecycle governance, contextual risk mapping, repeatable measurement and production controls as design principles for trustworthy AI operations.
NIST GenAI ProfileGenerative AI risks need specific controlsEvidence grounding, evaluation, monitoring, privacy controls and human review address risks that are specific to generative systems.
ISO/IEC 42001:2023AI management and continual improvementThe Core is designed around traceability, defined controls, monitoring and managed improvement of AI-enabled product behavior.
ISO/IEC 23894:2023AI-specific risk managementRisk controls are integrated into the design, deployment and operation of AI-assisted features instead of being treated as a one-time release check.
EU AI Act transparencyClear AI interaction and human oversightThe public architecture, identified AI features and explicit decision boundaries support a risk-based transparency approach while product-specific obligations remain subject to legal assessment.
OpenAI evaluation guidanceEval-driven developmentTask-specific datasets, automated graders, production-like cases and human judgment create a continuous quality process for non-deterministic systems.
Public architecture boundary
Transparent about controls. Private about the mechanism.
Collty publishes the principles, safeguards and accountability model needed to understand how AI is used. It does not publish proprietary ranking formulas, feature definitions, weights, thresholds, system prompts, private evaluation datasets, provider routing rules, encryption details or internal data schemas.
Back to About