Collty Intelligence Core connects your business model, project memory, people, AI agents and external tools. Shared context, domain intelligence and execution controls turn company knowledge into accountable work.
Explore how it worksOne intelligence foundation.
The context of your whole company.
Select relevant project facts with their sources, status and version.
The Intelligence Core is the middle layer: it turns operational data into context and grounded decisions. The Collty Operating System uses that intelligence in applications, processes and configured actions.
Goals, domains, processes, teams, projects and financial records form a connected operating model. The work has a purpose, an owner and relationships.
The Intelligence Core contains the business context model and Context Engine. It combines authorized context, memory, retrieval, domain algorithms and AI reasoning with evaluation and decision controls.
Business Architect, Agent Center, Team Intelligence and delivery tools use that foundation to build models, prepare decisions and execute configured actions.
Specialized reasoning, shared control. Models support interpretation and synthesis. Retrieval, access, domain logic, evaluation and the record of what happened stay connected through Collty.
See how company context becomes useful work. Choose a business need, follow its workflow and inspect the evidence, mechanisms and result at each stage.
Build a lead-to-client process with clear owners and handoffs.
Strategy & goals · Business model · Operating constraints
Bring the company's purpose, products, customers and constraints into the same design context.
Goals, constraints and ownership give AI a company-specific basis for the proposed model.
Domain structure organizes management, core and support work into a connected operating model.
Supported published flows refer to a reviewed model version and record their execution events.
Decisions, working rules and observations keep their source, version and state. The next plan, report or agent task can use the relevant knowledge within the permissions of the work.
Project goals, recorded decisions, task changes and accepted outcomes — linked to the project and their source versions. Earlier facts stay distinguishable from the current plan.
Select a memory layer to inspect what it holds and how it works. Select a use on the right to see which records can support that workflow and what still requires human authority.
Knowledge belongs to its company. A saved observation does not approve a decision or grant access.
People, agents, projects and processes refer to connected work objects. Coherence means preserving their identity, source version, state and authority as context moves from a record into a decision or action.
The project keeps its identity; the change records a new source version.
Earlier targets remain history. Draft alternatives retain their proposed status.
A report or agent task can refer to the confirmed version without treating the old target as today's plan.
A company owner can inspect what is recorded, where it came from and where saved knowledge was used. Privacy & Data in My Account brings the company memory map, export and lifecycle controls together.
Grouped sources show the amount of saved knowledge. Open a source to inspect its versions and recorded use, with historical gaps kept visible.
Pause memory collection, export a readable knowledge ZIP or choose an explicit deletion scope. Controls are restricted to the authorized company owner.
Company isolation and server-side authorization protect retrieval, export and actions. AI provider context is minimized; credentials and unrelated records are excluded.
Memory payloads are encrypted. Identities, permissions, statuses, timestamps and counters remain operational metadata. An exported ZIP is readable company data and needs protection at its destination.
Access checks, data handling and execution controls work together. These defenses reduce exposure to prompt injection and excessive agent authority; security still requires ongoing testing and operational review.
Server-side checks bind knowledge and operations to the authorized company, project and actor. Source access and enabled tool permissions limit the available scope; selecting a company or adding an AI instruction cannot grant new authority.
Documents, external records, tool descriptions and earlier outputs are treated as evidence, not authority to rewrite the workflow. Context filtering removes unrelated fields and redacts secrets. Fixed inputs and schema validation add executor-side controls alongside model instructions.
Connected writes prepare a reviewable operation before execution. Approval is tied to its frozen arguments and connection revision. Changing the connection requires a new approval, and fixed recipients cannot be replaced by generated arguments.
Connection transport checks endpoints and resolved addresses, blocks private and internal network destinations and pins the validated address for the request. HTTPS, response limits, timeouts and connection revocation support controlled access to external tools.
Request limits, bounded calls and execution deduplication reduce abuse and repeated effects. Paused or stopped workflow executions cannot be requeued by a stale worker result. Stopping later work does not undo an operation already completed by an external service.
Recorded runs and source versions preserve evidence and operation status. OpenTelemetry spans link execution identifiers while excluding prompts, outputs, source content and credentials. Protected incident procedures support investigation and corrective action.
Team assembly remains a dedicated intelligence system within the Core. It connects relevant professional evidence to the request, then helps managers review how roles, competencies and capacity cover the work together.
Only relevant, permitted professional evidence belongs in the assessment. Missing records remain visible without being treated as low ability. The specialist inspector shows facts, not an overall person score; the manager reviews profiles and chooses the team.
The project need is normalized into a task-specific semantic intent: outcome, scope, market, constraints and required capabilities.
Semantic retrieval finds the closest professional evidence by meaning, not by literal keyword overlap or a fixed profession dictionary.
Vector matches are resolved back to active, permitted Collty profiles and their source-backed competencies, experience, cases and work signals.
AI organizes relevant capabilities, linked cases and industry context for the requested role. Profile declarations are distinguished from project records; absent evidence is not evidence of low ability.
Team indicators summarize capability coverage, role balance, availability, allocations and collaboration context. They support review of the composition, not a guarantee of delivery or an individual quality rating.
Ranked team variants contain only real professionals and evidence-linked roles. Unsupported matches are rejected rather than invented.
Eligible outcomes become learning signals. A candidate version is evaluated against a baseline, tested on unseen evidence and observed before activation. People review the evidence and keep control of the release.
Run versioned private datasets through task-specific graders for assembly, planning and recommendations.
Feedback alone cannot change the active ranking. Evaluation, controlled rollout and a recorded review protect the release; the previous version remains available for rollback.
Access to information, permission to act and approval of a consequential change are separate controls. The workflow stays tied to the authorized user, relevant evidence and the configured action.
Task-relevant context is selected and filtered. Credentials, payment secrets and unrelated workspace records do not belong in the reasoning packet.
Data & privacyAn agent’s role describes its work. Internal permissions, connected-account access, action settings and approval rules determine what it can actually do.
Instructions are not access grants.Execution results and relevant provenance connect decisions to the task, evidence and model versions. Proposed, applied, blocked and failed work remain distinguishable.
A prepared draft is not a completed action.Each AI workflow has its own intended purpose, classification, data boundary and oversight rules. Team recommendations and task allocation require controls appropriate to their effect on people.
Intended purpose, owner, Collty's provider role, the customer's deployer role, risk class, human-oversight boundary and review date are versioned per product system. Team Assembly, Signals applicant ranking, Team Intelligence and task allocation are conservatively governed as Annex III high-risk systems.
The risk register covers discrimination, ranking and selection bias, false exclusion, unsupported inclusion and misuse. Versioned data contracts define allowed evidence, protected-attribute exclusions, proxy tests, missing-evidence behavior and activation gates.
A displayed recommendation can explain the professional evidence that supported it. A specialist can correct their profile or report an AI-ranking error; the customer still makes and owns the commercial team decision. Hidden search candidates are not disclosed.
Quality, false inclusion and exclusion, overrides, review requests, drift, reliability and misuse signals feed a protected incident and corrective-action process. Role-specific AI literacy records support trained human oversight.
Ratings enter the gated learning loop. A system-error or evidence-correction request protects the accuracy of Collty’s output. Collty does not adjudicate the customer’s team choice or expose hidden search candidates to manufacture an explanation.
Profile guidance can help professionals make their evidence clearer; it does not guarantee selection. The customer or project owner makes the final team choice.
Read the AI termsTechnical specifications define how provenance is exported and AI operations are traced. NIST and ISO guide the risk controls and operating procedures. These implemented mechanisms are distinct from a completed conformity assessment or certification.
Govern: versioned system inventory, responsibilities and risk contracts. Map: intended purpose, authorized context and affected people. Measure: task-specific evaluations, holdout comparisons and production monitoring. Manage: activation gates, human approval, incident controls and rollback. Completed organizational reviews require their own recorded evidence.
Official sourceSource grounding, exact memory-version references, minimized provider context, protected-attribute exclusions, versioned evaluations and human oversight address generative AI risks. Risk treatments are defined per workflow; an AI response is recorded separately from an approved business decision.
Official sourceCollty's AI management procedures define system scope, accountable owners, risk and impact review, release evidence, monitoring, incidents and corrective improvement. Technical gates support these procedures. Training, internal audit and management review require actual human participation and records; ISO certification has not been claimed.
Official sourceAI-specific risk management connects purpose and context to named risk owners, selected treatments, measurable acceptance criteria and review triggers. Changes to sources, providers or action authority reopen the assessment, while evaluations, approval and recovery evidence support lifecycle decisions.
Official sourceOwner-authorized company knowledge exports include PROV-O JSON-LD. Entities, activities and agents connect saved source versions to their recorded origin, revisions and use in AI operations. Exact memory references preserve which version was used; missing historical references remain unknown.
Official sourceThe same company export includes PROV-JSON, checked against its published schema. It preserves recorded generation, derivation, revision and use relationships for other tools. PROV-JSON is a W3C Member Submission; PROV-O is a W3C Recommendation.
Official sourceThe official Node SDK creates spans for Core AI operations, retains parent trace context and links trace and span identifiers to recorded runs. Provider, model, operation and success or error status support diagnosis without exposing prompts, responses, source content or credentials. OTLP export requires a configured collector; external collection is not enabled by the SDK alone.
Official sourceThe public architecture, identified AI features and explicit decision boundaries support a risk-based transparency approach while product-specific obligations remain subject to legal assessment.
Official sourceTask-specific datasets, automated graders, production-like cases and human judgment create a continuous quality process for non-deterministic systems.
Official sourceExplore the architecture, data boundaries and operating controls in more detail.
It is Collty's intelligence architecture for the connected business: strategy, processes, people, agents, projects, economics and external tools. It connects the permission-aware work graph, semantic retrieval, domain algorithms, evaluation, model routing, provenance and controlled learning. External reasoning models are replaceable components inside this system, not the owner of its evidence or decision policy.
No. It is a domain intelligence system built around Collty's own evidence model and algorithms. Specialized models contribute interpretation and synthesis where they add value; retrieval, ranking, permissions, validation, evaluation and activation controls remain inside Collty.
Only authorized, task-relevant professional evidence is considered: competencies, roles, experience, cases, industries, languages, availability, capacity, delivery history, verified outcomes and collaboration signals. Team-level analysis also examines coverage, complementarity, workload, resilience and commercial context when the project requires it.
No. It shows listed capabilities and cases, reported availability, confirmed allocation hours and collaboration evidence with source categories. Missing evidence stays unknown rather than receiving a default rating. Shared competencies or industry declarations are separate from recorded shared work. The manager checks relevance with the specialist and makes the assignment decision.
Project pulse is a voluntary check-in about how a project is going. It appears in Project management and is separate from task status, Team intelligence and specialist assessments. One latest response per participant contributes for 30 days; a summary appears after three responses. When support is needed, a short questionnaire helps prepare an editable private message to a project manager. Reasons and details stay in the draft until you send it. No individual quality badge is awarded.
The assessment is designed around professional evidence and derived work signals, not sensitive personal attributes. Credentials, payment details, encryption material and unrelated workspace content are excluded. Context is minimized and filtered before any provider request.
Collty does not send customer workspace content to model providers for training shared foundation models. Provider processing is used to deliver the requested feature under the applicable service terms, while Collty keeps its work graph, evaluation data and learning controls within its own platform boundary.
The Core builds a task-specific evidence packet instead of exposing the work graph directly. Allowlisted structures, field filtering, redaction and evidence identifiers reduce the context to what the current task requires. Deeper protected traces are retained only when needed for evaluation, security or investigation.
Generated conclusions must resolve to permitted source objects and grounded evidence identifiers. Domain graders detect unsupported members, invented evidence and missing mandatory capabilities, preventing fluent model output from being accepted without source-backed support.
Explore how Collty connects strategy, processes, people, agents and economics in one operating model.
Explore the Collty platformThis page explains the public architecture and its controls. Proprietary ranking formulas, prompts, weights, private evaluation datasets, routing policies and internal schemas remain private.