Innovation Child
PLATFORM ARCHITECTURE

From Raw Data
to Delivered Insight

Four composable layers. Each builds on the one below. No magic — just architecture.

01
01 · FOUNDATION

The Data Layer

Four agent-compliant components — ingestion, retrieval, property data, and extensible custom tools. Anything IC Agents can call lives here.

IC EXTRACTOR
AGENT COMPLIANT

Multi-format document ingestion normalized to a consistent JSON schema.

  • PDF, images, video, audio, and structured tables — single pipeline
  • Extracts charts, diagrams, and complex multi-column layouts at ≥99% fidelity
  • Unified schema — downstream agents never parse format-specific quirks
VS VANILLAPyPDF2 / pdfplumber are text-only; Tesseract needs per-format wiring; no pipeline produces a shared schema.
IC VECTORDB
AGENT COMPLIANT

Semantic search with multilingual embeddings, deduplication, and reranking.

  • BGE-M3 outperforms OpenAI ada-002 on multilingual and domain-specific recall
  • MMR deduplication removes near-duplicate chunks before they reach the model
  • Cross-encoder reranking lifts top-k precision without re-embedding
VS VANILLAPlain Qdrant / Pinecone with cosine similarity returns duplicates and ranks by distance, not relevance.
IC PROPERTY
AGENT COMPLIANT

Planning zones, council regulations, school catchments, and suburb data — unified.

  • Planning scheme + overlays + schools + suburb profiles in one cross-referenced feed
  • Interprets raw planning data into plain-language restrictions and action implications
  • Victorian end-to-end — no stitching CoreLogic, Domain, and council PDFs separately
VS VANILLACoreLogic and Domain expose sales data only; planning info is buried in council PDFs; schools need a separate ACARA feed.
+ EXTENSIBLE
EXTENSIBLE

Any compliant data source or tool can be registered and called by IC Agents.

  • Standardized skill interface — implement the spec, plug in, done
  • No lock-in: REST endpoint, Python function, or MCP server
  • Scoped to the agent identity that needs it — not globally exposed
VS VANILLALangChain tools require boilerplate wrappers, lose typing, and are exposed to all agents by default.
orchestrated by
02
02 · ORCHESTRATION

IC Agents

The harness that dispatches skills and tools to the foundation. Token-efficient, deterministic, identity-driven.

ORCHESTRATION ENGINE

IC Agents

Token-efficient, deterministic agentic harness. Dispatches skills and tools to foundation components — each agent operates with a defined identity, knowledge scope, and tool manifest.

  • LangGraph-based plan → execute → verify loop — deterministic, not probabilistic
  • Each agent carries a defined identity, knowledge scope, and tool manifest
  • Structured prompt architecture cuts token cost 40–60% vs LangChain defaults
  • Multi-agent coordination with shared state and interrupt handling

REGISTERED TOOLS

IC EXTRACTOR
IC VECTORDB
IC PROPERTY
+ EXTENSIBLE
VIA STANDARDIZED SKILL INTERFACE
powers
03
03 · INTELLIGENCE

Thinking Models

Expert reasoning and quantitative decision engines — trained from industry leaders, used to generate insights.

IC THINKERS

EXPERT THINKING MODELS

IC Thinkers

Multi-persona reasoning engine — IC-distilled expert knowledge, structured debate, and cross-verification.

Trained from documented reasoning patterns, investment theses, and public analysis — not role-playing, but genuine knowledge distillation.

Multi-personaCross-verificationDeep research
PANEL DECISIONCOMING SOON

EXPERT + QUANTITATIVE FUSION

Panel Decision

Structured panel deliberation engine for complex, high-stakes enterprise decisions.

An expert panel deliberates; a quantitative model scores. Two signals fused into one auditable recommendation with full reasoning trace.

Weighted deliberationQuantitative scoringAudit trail
delivered as