Core Capabilities

Foundational primitives available in every system we build.

Reasoning & Planning

Multi-step reasoning, task decomposition, adaptive planning, and self-reflection. Supports ReAct, Plan-and-Execute, and custom planning strategies.

  • Chain-of-thought reasoning
  • Dynamic plan generation
  • Self-correction loops
  • Multi-agent coordination
Technical details

Tool Use & Function Calling

Typed, permissioned tool invocations with schema validation, retries, timeouts, and human approval gates. 40+ pre-built tools.

  • JSON Schema validation
  • Parallel & sequential execution
  • Idempotency for mutations
  • Circuit breakers & retries
Technical details

Memory Systems

Four-tier memory architecture: working, episodic, semantic, and procedural. Scoped per user/team/org with privacy guarantees.

  • Working: session context
  • Episodic: conversation history
  • Semantic: facts & preferences
  • Procedural: skills & workflows
Technical details

Retrieval & RAG

Hybrid search (vector + keyword + graph) with cross-encoder reranking, citation enforcement, and access-controlled knowledge bases.

  • Semantic + lexical fusion
  • Cross-encoder reranking
  • Citation verification
  • Document-level ACLs
Technical details

Model Routing

Intelligent task-to-model mapping based on complexity, latency budget, cost ceiling, data sensitivity, and evaluation results.

  • Frontier for reasoning
  • Efficient for classification
  • Local for sensitive data
  • Cost-optimised routing
Technical details

Structured Output

Guaranteed schema compliance via constrained generation (JSON Schema, function calling, guided grammar). No parsing failures.

  • JSON Schema enforcement
  • Type-safe tool results
  • Validation at generation
  • Retry on format error
Technical details

Quality & Operations

Capabilities that make systems production-ready.

Evaluation Framework

Automated eval suites: accuracy, reliability, latency, cost, hallucination, instruction following, tool reliability, edge cases, regression, security.

  • Golden datasets as code
  • CI-integrated regression
  • Production sampling
  • Human + LLM judges
Technical details

Guardrails

Defence in depth: input (PII, injection, topics), model (constraints, format), output (citations, validation), tool (permissions, approvals), infrastructure (network, secrets).

  • Input classification
  • Output verification
  • Tool permission gates
  • Audit logging
Technical details

Observability

Full-stack: traces (request flow), metrics (RED + quality), logs (structured, correlated), automated eval sampling, drift detection.

  • Distributed tracing
  • Quality dashboards
  • Cost tracking
  • Anomaly alerts
Technical details

Deployment & Rollout

Progressive rollout (canary → ramp), feature flags per capability, instant rollback (<10s), blue/green, immutable deployments.

  • Canary with auto-rollback
  • Feature flags everywhere
  • Config as code
  • Multi-region DR
Technical details

Human-in-the-Loop

Approval gates, clarification requests, review queues, escalation paths. Slack/Teams/Email notifications with structured actions.

  • Risk-based approval
  • Batch review queues
  • SLA tracking
  • Custom integrations
Technical details

Security Controls

Threat modelling (STRIDE), penetration testing, red-teaming (injection, extraction, hijacking), compliance mapping (SOC 2, GDPR, ISO 27001 conceptual).

  • Data classification
  • Tokenisation for PII
  • Zero-retention APIs
  • Supply chain scanning
Technical details

Model Support

We don't marry a single provider. Models are selected per task via our router.

Tier Models Use Cases Deployment
Frontier GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro Complex reasoning, planning, coding, synthesis API (managed)
Specialised CodeLlama, Nemotron, DeepSeek-Coder Code generation, structured output, SQL API / Self-hosted
Efficient Llama 3.1 70B, Phi-3, Gemma 2, Qwen 2.5 Classification, extraction, routing, local deploy Self-hosted (vLLM/TGI)
Embedding text-embedding-3-large, BGE-M3, Nomic, E5 Retrieval, clustering, similarity, reranking API / Self-hosted
Multimodal GPT-4o, Claude 3.5, Gemini 1.5 Vision, audio, document understanding API (managed)

Principle: Model routing selects the right model per task based on complexity, latency budget, cost ceiling, data sensitivity, and evaluation results. No vendor lock-in.

Want to see the architecture?

Explore the full system architecture with component diagrams and data flows.

View AI Architecture →