Loomra Anatomy: Five Layers Explained

Deep Dive | ~12 min read
Track: 02 Architecture and Technology
Module: 02.1 The Two-Layer Platform Architecture
Tag: Essential

Every enterprise HR platform makes promises about AI. Most bolt a language model onto a database and call it intelligent. Loomra is something fundamentally different: a purpose-built workforce intelligence stack, eight years in the making, trained on billions of real career events, and architected from the ground up for the specific complexities of talent decisions.

This article breaks down the five layers that make Loomra work — not as a marketing overview, but as a technical anatomy. If you are an IT architect evaluating Gloat, an HR technologist designing your talent stack, or simply someone who wants to understand what happens between the moment an employee asks “what’s my next career move?” and the moment they receive a personalized, policy-compliant answer, this is where you start.

The Design Philosophy: Why Five Layers?

Most AI systems in HR follow a monolithic pattern: data in, prediction out. That works for narrow tasks. It fails catastrophically when you need the system to reason across multiple domains simultaneously — matching a person to a role while respecting compensation bands, compliance policies, visa restrictions, and the employee’s own aspirations.

Loomra solves this with a layered architecture where each component has a single, well-defined responsibility:

The critical insight is that these layers are not sequential — they are composable. A single agent request might traverse all five layers in parallel, assembling a response that is simultaneously personalized, semantically accurate, organizationally compliant, and grounded in real entity relationships. No single layer could produce that result alone.

Layer 1: Knowledge Graph

The Knowledge Graph is the structural backbone of Loomra. While most HR platforms store data in flat relational tables — employees in one table, jobs in another, skills in a third — the Knowledge Graph represents the workforce as an interconnected network of entities and relationships.

Scale: A typical enterprise deployment contains 2.4 million entities connected by 18.7 million relationships. These are not static records. The graph updates continuously as new career events flow in from connected systems — Workday, SAP SuccessFactors, Oracle HCM, and other HCM platforms.

Eight relationship types form the grammar of the graph:

• People → Skills — What capabilities does each person hold, at what proficiency?
• Skills → Tasks — Which tasks require which skills, and at what level?
• Tasks → Roles — How do tasks compose into defined roles?
• Roles → People — Who holds or has held each role?
• Candidates → Roles — Which candidates are viable for which positions?
• Skills → Learning — What learning content develops which skills?
• Projects → Skills — Which skills does each project require or develop?
• Roles → Market — How does each role map to external labor market data?

Performance: The graph supports 6-hop queries — traversals that cross six relationship boundaries — in under 50 milliseconds. That means when the system needs to answer “which employees have skills adjacent to the skills required by a role that is similar to a role in a department that is planning a restructuring,” it can do so in real time. The graph operates at a 99.9% SLA.

Why this matters: A 6-hop query is what separates genuine workforce intelligence from keyword matching. When a Succession Planning Agent needs to identify internal candidates for a VP of Engineering role, it does not search for people with “VP” in their title. It traverses from the role to its required skills, from those skills to adjacent skills, from adjacent skills to people who hold them, from those people to their career trajectories, and from those trajectories to readiness signals — all in one query, all in under 50 milliseconds.

Layer 2: Intelligent Tools

Intelligent Tools are the operational capabilities that Loomra exposes to agents. Think of them as the verbs in the system’s vocabulary — the actions it can actually perform. There are 14 purpose-built tools organized across five categories:

Retrieve and Understand

• Semantic Retriever — Finds contextually relevant information across the entire talent data lake using meaning, not keywords.
• Entity Retriever — Resolves specific entities (a person, a role, a skill) from partial or ambiguous references.

Match and Recommend

• Personalized Matching Engine — Produces multi-factor matches between people and opportunities, weighing skills, aspirations, readiness, and organizational need.
• Role Progression Router — Maps viable career paths from any current role, based on historical transition patterns across millions of career events.

Predict and Model

• Predictive Attrition Modeler — Identifies flight risk signals by correlating engagement patterns, career velocity, market conditions, and peer benchmarks.
• AI Disruption Modeler — Projects which roles and skills face disruption from automation and AI adoption, with timeline estimates.
• Future-State Readiness Index — Scores organizational readiness for strategic workforce shifts (M&A integration, market expansion, technology transformation).

Act and Automate

• Profile Builder — Constructs and enriches employee profiles by synthesizing data from multiple systems into a unified talent record.
• Trigger Agent — Monitors conditions across the graph and fires actions when thresholds are met (e.g., “notify manager when three direct reports show declining engagement”).
• Entity Generator — Creates new entities in the graph: new roles, new skill definitions, new project structures.
• Entity Updater — Modifies existing entities based on new data or organizational changes.
• Talent Pipeline — Builds and maintains candidate pools for current and anticipated openings.

Govern and Report

• Algorithmic Equity Layer — Continuously audits tool outputs for bias across protected categories, flagging and correcting disparate impact before results reach users.
• Report Generator — Produces structured analytics and narrative summaries from graph queries and tool outputs.

Performance: Collectively, Loomra’s Intelligent Tools have delivered over 200 million matches across production deployments, with an average response time of under 100 milliseconds.

Why this matters: General-purpose AI can generate plausible-sounding HR recommendations. Loomra’s tools generate accurate ones. Each tool encodes domain-specific logic that took years to develop — the Predictive Attrition Modeler, for instance, does not simply flag employees who update their LinkedIn profiles. It correlates dozens of internal signals (project completion patterns, skill development velocity, peer network changes, manager effectiveness scores) against labor market conditions specific to that employee’s role and geography.

Layer 3: Personalization Engine

The Personalization Engine is what makes Loomra feel like a personal career advisor rather than a corporate system. Trained on over 5 million employee career patterns, it builds and maintains a living model of each individual employee.

The engine synthesizes data from every connected system — HCM platforms (Workday, SAP SuccessFactors, Oracle HCM), learning management systems, applicant tracking systems, and collaboration tools — into a unified profile that captures four dimensions:

1. Skills — Not just self-reported or manager-assessed skills, but inferred skills based on project participation, learning completion, peer interactions, and output patterns.
2. Aspirations — Where the employee wants to go, captured through explicit goal-setting, implicit signals (what content they consume, what roles they explore), and career conversation data.
3. Collaboration patterns — Who they work with, how they contribute to teams, their network breadth and depth within the organization.
4. Career momentum — The velocity and direction of their career trajectory. Are they accelerating, plateauing, or at an inflection point?

The result is coach-quality career guidance delivered at enterprise scale. A large organization might have 50 career coaches for 50,000 employees. The Personalization Engine provides individualized guidance to every single one of them.

Why this matters: When a Growth Agent recommends a stretch assignment to an employee, that recommendation is not based solely on skill match. The Personalization Engine factors in the employee’s stated aspirations, their demonstrated learning velocity (can they realistically close the skill gap?), their collaboration network (do they have relationships in the target team?), and their career momentum (is this the right time for a lateral move?). The same stretch assignment might be recommended to one employee and explicitly withheld from another — not because of skill differences, but because the Personalization Engine understands they are at different career moments.

Layer 4: Retrieval and Embedding

Retrieval and Embedding is the layer that makes Loomra understand meaning — the semantic intelligence that lets the system know that “people management” and “team leadership” refer to overlapping capabilities, even though they share no words in common.

Loomra runs six proprietary embedding models, each purpose-built for a specific workforce AI task:

The critical performance metric: Loomra achieves 90% matching accuracy on workforce tasks. General-purpose embedding models (the kind you get from OpenAI or open-source repositories) score between 56% and 66% on the same benchmarks. That 24-34 percentage point gap is the difference between a system that makes useful recommendations and one that generates noise.

Why the gap is so large: General-purpose models are trained on internet text. They know that “Java” is a programming language and an island. They do not know that a “Senior Java Developer” and a “Staff Backend Engineer” are often the same role, that “stakeholder management” is a stronger signal for leadership readiness than “project management,” or that “Python” in a data science context implies different adjacent skills than “Python” in a DevOps context. Loomra’s models encode these distinctions because they were trained on real career transitions, real skill adjacencies, and real hiring outcomes — over a billion of them.

The skill harmonization model (gloat-harmonize-skill) deserves special attention: it operates across 19 languages, which means a skills taxonomy created in English automatically maps to equivalent skills described in German, Japanese, Portuguese, or Mandarin. For global enterprises operating across 112 countries, this eliminates the taxonomy fragmentation that plagues most skills-based organizations.

Layer 5: Business Logic Engine

The Business Logic Engine is the layer that most AI vendors ignore and most enterprises need most. It ensures that every output from Loomra — every match, every recommendation, every agent action — complies with the organization’s specific policies, rules, and constraints.

The engine operates across 24 rule categories, extracting and enforcing constraints from:

• Approval workflows — Who must sign off before a transfer, promotion, or gig assignment?
• Role-based access control (RBAC) — Which managers can see which talent pools? Which employees can access which opportunities?
• Eligibility criteria — Minimum tenure requirements, performance thresholds, certification prerequisites.
• Organizational hierarchy — Reporting structures, span-of-control limits, cross-functional movement policies.
• Compensation bands — Salary ranges, equity guidelines, geographic pay differentials.
• Compliance policies — Labor law requirements, visa restrictions, union agreements, equal opportunity mandates.

The three-step process:

1. Extract — Rules are ingested from source systems (HRIS configuration, policy documents, approval matrices) and from explicit administrator input.
2. Normalize — Rules are translated into a standard enforcement format, resolving conflicts and establishing precedence hierarchies.
3. Enforce — Every Loomra output passes through the rule engine before reaching users. Non-compliant results are filtered, flagged, or modified.

Performance: Rule evaluation completes in under 15 milliseconds per query — fast enough that compliance checking adds no perceptible latency to user interactions.

Certifications: The Business Logic Engine underpins Gloat’s SOC 2 Type II, ISO 27001, and GDPR compliance certifications. Audit trails are maintained for every decision, every rule evaluation, and every agent action.

Why this matters: Consider a Retention Agent that identifies an at-risk senior engineer and recommends a counter-offer involving a promotion and equity grant. Without the Business Logic Engine, that recommendation might violate the employee’s compensation band, skip a required approval step, or conflict with a pending reorganization. The Business Logic Engine catches all of this in 15 milliseconds, before the recommendation ever reaches a manager’s screen. The agent does not need to know every rule — it simply passes its proposed action through the engine and receives a compliant version back.

How Agents Use Loomra

Gloat deploys 29 purpose-built agents across four categories:

• Talent Management — 7 agents (retention, growth, succession, engagement, performance, internal mobility, mentoring)
• Talent Acquisition — 6 agents (sourcing, screening, scheduling, offer optimization, onboarding, employer branding)
• Skills — 8 agents (skills inference, gap analysis, taxonomy management, learning recommendation, certification tracking, skills validation, market benchmarking, skills architecture)
• Workforce Planning — 8 agents (headcount planning, scenario modeling, org design, cost optimization, contingent workforce, location strategy, M&A integration, strategic workforce planning)

Each agent is a specialized orchestrator that combines Loomra’s five layers to execute a specific talent workflow. Here is how a single agent request flows through the stack:

Example: Retention Agent detects flight risk

1. The Personalization Engine detects declining career momentum for a senior product manager — engagement scores are flat, skill development has stalled, and the employee has been browsing internal roles without applying.

2. The Knowledge Graph is queried to understand the employee’s full context: their skill profile, team relationships, project history, manager effectiveness score, and the market demand for their role in their geography.

3. The Retrieval & Embedding layer identifies semantically similar employees who left the organization in the past 18 months, surfacing the patterns that preceded their departures.

4. The Intelligent Tools layer runs the Predictive Attrition Modeler, which synthesizes all of this into a calibrated risk score and generates three retention intervention options: a stretch project, a mentorship pairing, and a compensation adjustment.

5. The Business Logic Engine evaluates each intervention against organizational policies. The compensation adjustment exceeds the mid-cycle adjustment cap — it is flagged for VP approval. The stretch project requires a skill the employee has not yet validated — a learning prerequisite is attached. The mentorship pairing clears all rules.

6. The Retention Agent delivers the compliant recommendations to the employee’s manager through their preferred delivery surface — Microsoft Teams, Slack, Google Chat, or Copilot — with full context on why each option was recommended and what approvals are needed.

Total elapsed time from signal detection to manager notification: seconds, not days.

This is possible because Loomra is model-agnostic. The orchestration layer works with Anthropic, Google, IBM WatsonX, or your own infrastructure. The proprietary intelligence — the Knowledge Graph, the embedding models, the business rules — lives in Loomra’s layers, not in any single LLM. The language model is a reasoning surface; Loomra is the intelligence.

The Layers in Concert

The real power of Loomra’s architecture becomes visible when you trace how the five layers interact across different scenarios:

Scenario: Organization announces a restructuring affecting 200 roles

• The Knowledge Graph maps every affected employee’s skills, relationships, and role dependencies.
• The Retrieval & Embedding layer identifies the closest-match internal roles for each displaced employee, across all business units and geographies.
• The Personalization Engine ranks those matches by individual fit — not just skill overlap, but aspiration alignment, learning velocity, and career momentum.
• The Intelligent Tools layer runs the Future-State Readiness Index to project which receiving teams can absorb new members and which skill gaps the redeployments would create or close.
• The Business Logic Engine filters every proposed move through compensation policies, visa restrictions, union agreements, and approval workflows.

The result: a complete, policy-compliant redeployment plan for 200 employees, generated in minutes instead of the weeks it would take a team of HR business partners working with spreadsheets.

Scenario: New market expansion requiring 150 net-new hires across 3 countries

• The Knowledge Graph identifies which skills already exist internally and which must be hired externally.
• The Retrieval & Embedding layer benchmarks internal skill descriptions against local market terminology in each country’s language.
• The Intelligent Tools layer builds talent pipelines for external hires while the Role Progression Router identifies internal employees who could relocate or transition.
• The Personalization Engine assesses which internal candidates would actually want an international move, based on their stated aspirations and demonstrated preferences.
• The Business Logic Engine validates every proposed hire and transfer against local labor law, immigration requirements, and compensation bands for each geography.

What to Read Next

Each of Loomra’s five layers warrants its own deep dive. Continue your exploration with these articles:

• The Knowledge Graph in Depth — Entity types, relationship modeling, query optimization, and how the graph evolves over time.
• Intelligent Tools Reference — Detailed specifications for all 14 tools, including input/output schemas, accuracy benchmarks, and integration patterns.
• Personalization at Scale — How the Personalization Engine builds and maintains profiles for millions of employees without manual data entry.
• Embedding Models for Workforce AI — Architecture details on all six proprietary models, training methodology, and benchmark comparisons against general-purpose alternatives.
• Business Logic and Governance — Rule category reference, policy extraction workflows, and compliance certification details.
• The 29 Agents — Complete catalog of Gloat’s agent library, organized by category, with workflow diagrams and deployment guides.

Loomra powers Gloat’s Agentic HR platform across 100+ enterprise customers, 5M+ employees, 112 countries, and 200+ deployments. It is the product of eight years of dedicated workforce AI research — purpose-built, not bolted on.