Instead of one‑off bots, we use a repeatable seven‑layer stack so each new pilot is faster and safer to build.
Experience: chat widgets, WhatsApp, email, web, and internal portals.
Orchestrator: routes tasks, calls agents, applies policies, and escalates to humans.
Agents: specialised workers for triage, RAG Q&A, quoting, CX, vision, and DataOps.
Intelligence: language models, embeddings, retrieval and re‑ranking, and domain prompts.
Tools & Connectors: ERP, CRM, databases, file stores, email, WhatsApp API, OCR, and vision models.
Data: mini‑data lake with bronze/silver/gold layers and vector indices.
Observability & Security: logging, metrics, cost and latency tracking, RBAC, PDPL, and guardrails.
We keep a standard catalog of agents that we configure per client, instead of inventing a new bot for each project.
Factory manager: top‑level orchestrator that routes requests and enforces policies.
Triage/NLU agent: detects intent and picks the right tools and agents.
RAG answering agent: retrieves context, generates cited answers, and falls back to “I don’t know” when needed.
Quoter agent: reads RFQs, maps to prices and policies, and drafts quotes.
CX agent: responds to customers across channels with safe templates and fallbacks.
Vision agent: runs simple quality or safety checks on images or video frames.
DataOps and EvalOps agents: handle ingestion, quality checks, and evaluation suites.
We structure data into three tiers, so agents always work from the right level of detail and quality.
Bronze: raw ingested data from ERP, files, email exports, and logs; immutable and auditable.
Silver: cleaned, typed tables and normalised documents ready for agents and dashboards.
Gold: curated KPIs, trusted knowledge articles, and up‑to‑date embeddings and indices.
Automatic checks for freshness, quality, and drift across layers.
Data catalog and lineage so IT and business teams know what powers which agent.
The AI Factory is built with guardrails and governance baked in. We treat PDPL and your internal policies as core design inputs—not optional extras.
Content and relevance filters to keep agents within defined topics and scopes.
PII redaction and masking where appropriate, with minimal exposure in logs and traces.
Role‑based access to data sources and tools, with full audit trails of actions.
Read‑only first, then gated writes with approval workflows and rollback plans.
Human‑in‑the‑loop portals for reviewing low‑confidence or sensitive actions before execution.
Once the first pilot is successful, scaling is about adding new agents and workflows on the same factory—rather than starting from scratch.
Reuse connectors, prompts, evaluation sets, and guardrails across pilots.
Extend pilots horizontally (more use‑cases) or vertically (more depth, channels, or automation).
Transition from project mode to a factory operating model with clear ownership.
Integrate AI Factory metrics into your existing BI and governance dashboards.
Keep options open: mix cloud LLMs with in‑Kingdom or on‑prem models as your needs evolve.
For IT, security, and data teams, we offer focused sessions to walk through the AI Factory architecture, discuss constraints, and align on a safe delivery plan.
Review the seven layers in the context of your systems and policies.
Map candidate pilots to existing data sources and channels.
Agree on deployment patterns: SaaS, VPC in KSA, or on‑prem.
Define responsibilities between LeenAI, your IT team, and your security/compliance functions.