intelligence-engine-showcase

Intelligence Engine

A schema-driven intelligence engine that turns structured source material into searchable knowledge graphs, hybrid retrieval and an interactive AI-assisted application.

Public showcase: fbratten.github.io/intelligence-engine-showcase

Current evidence status

This recruiter-facing refresh was verified on 2026-08-02 against the private source repository at:

fbratten/intelligence-engine
main @ 3cdf7099c32c614a376a03324f545cfe8bc3a8b4

Current source-supported surface	Count or status
Product domains	3: code, archaeology, written_content
Supported source languages	8
MCP tools	14
REST endpoints	46
Source-reported tests	1261+
Backend	Python + FastAPI
Frontend	React + TypeScript + Sigma.js
Graph and vector storage	KuzuDB + NetworkX + LanceDB

The test figure is a source-reported claim at the pinned commit. This public showcase refresh did not re-run the private suite. See the full current evidence record for evidence classes, limitations and freshness rules.

The earlier longer README has been preserved at archived/README.pre-recruiter-refresh-2026-08-02.md.

---

The problem it solves

Large codebases, document collections and specialist datasets are difficult to inspect through one search method alone.

Keyword search misses semantic similarity. Vector search can lose structural context. Graph traversal is powerful but requires the right schema and entry point. Intelligence Engine combines the three and exposes them through a web application, REST API and Model Context Protocol server.

structured source
-> domain extractor
-> domain-scoped entities and relationships
-> KuzuDB + LanceDB + BM25
-> hybrid retrieval and graph expansion
-> FastAPI / MCP
-> React graph explorer, dashboards and AI-assisted inspection

What was built

• A schema-driven domain layer where YAML definitions control entities, relationships, properties, search profiles and UI behavior.
• Tree-sitter extraction for Python, JavaScript, TypeScript/TSX, Java, Go, HTML and CSS.
• Custom extraction for non-code domains.
• KuzuDB graph storage with NetworkX as an alternate backend and parity oracle.
• LanceDB semantic retrieval and BM25 lexical retrieval.
• Reciprocal-rank fusion across lexical, semantic and graph-context results.
• A FastAPI application surface with project, graph, search, audit, AI and operational endpoints.
• A React/TypeScript/Sigma.js graph explorer with search, filtering, entity detail, Cypher inspection, dashboards and AI data views.
• A FastMCP server exposing project indexing, retrieval, context, quality, health, Cypher, summaries, cross-project analysis and AI memory.
• Multi-project and shared-storage modes with project scoping, migration and collision controls.
• Incremental indexing with preservation of persistent AI overlays.
• Provider-aware AI summaries and Q&A with explicit locality and persistence reporting.

---

Product domains

1. Code intelligence

The built-in code domain extracts functions, classes, methods, modules, variables, interfaces, templates, forms, selectors and other language-specific structures.

Representative relationships include:

• CALLS
• IMPORTS
• EXTENDS
• DEFINES
• METHOD_OF
• LINKS_STYLESHEET
• REFERENCES_SCRIPT
• USES_VARIABLE

Typical uses:

• dependency and blast-radius inspection;
• cross-project search;
• dead-code, coupling and cycle analysis;
• source-grounded AI summaries;
• change-risk orientation before editing.

2. Archaeology

The archaeology MVP validates that the engine is not hard-coded to source code.

Its domain schema models:

• finds and artefacts;
• sites and locations;
• periods;
• materials;
• water bodies;
• relationships such as FOUND_AT, DATED_TO, MADE_OF, NEAR_WATER and ASSOCIATED_WITH.

3. Written content

The written-content domain handles long-form material such as books, manuscripts and articles.

It provides:

• structural extraction for documents, parts, chapters, sections and paragraphs;
• domain-specific views and evidence;
• bounded self-vs-self originality analysis;
• staged lexical and semantic audit processing;
• reviewable audit artifacts and comparisons.

Boundary: external-source plagiarism checking is planned. It is not claimed as complete.

Architecture fixture

A fourth schema, conformance_probe, exists solely to test generic domain rendering and capability handling. It is deliberately not user-indexable and is not counted as a product domain.

---

Screenshots

Knowledge Graph — interactive graph exploration	Cypher Console — bounded graph inspection
Dashboard Timeline — indexing history	Dashboard Health — structural signals
Cross-Project Compare	AI Memory — persistent AI data views

These screenshots are public evidence of earlier runtime captures. They are not represented as pixel-identical captures of the current private source pin.

---

Hybrid retrieval

The retrieval system combines:

1. BM25 lexical search for exact and near-exact terms;
2. semantic vector search for concept similarity;
3. graph-context expansion for related entities and structural evidence;
4. reciprocal-rank fusion to produce one ranked result set.

The right strategy depends on the question. The public search strategy demo explains when to prefer lexical, semantic, graph or hybrid retrieval.

---

Public MCP surface

The current private source exposes 14 MCP tools:

Tool	Purpose
ie_index	Index a project into a selected domain
ie_query	Search within one project
ie_search_all	Search across indexed projects
ie_context	Inspect entity context and relationships
ie_detect_changes	Orient change-risk analysis from Git state
ie_cypher	Run validated read-only Cypher queries
ie_wiki	Generate graph-grounded documentation
ie_status	List indexed projects and state
ie_health	Report structural health signals
ie_quality	Report code-quality metrics
ie_summarize	Generate a source-grounded entity summary
ie_batch_summarize	Summarize a project in batches
ie_global_analysis	Analyze shared multi-project graph state
ie_memory	Browse persistent AI-generated data

The source also exposes self-describing MCP resources and prompt workflows. Private project names, source content, filesystem paths, credentials and estate data are not included in this showcase.

---

Reliability and safety work

The project includes engineering beyond the happy path, including:

• read-only serving modes and bounded write surfaces;
• project-scoped shared-storage operations;
• collision prevention for indexing;
• preservation-first migration and reconciliation;
• refusal of unbindable Cypher placeholders on public query surfaces;
• explicit reporting when LLM inference succeeds but persistence does not;
• endpoint-locality controls for local and cloud providers;
• cancellation checks inside expensive audit stages;
• bounded audit inputs, outputs, workers and retained history;
• negative controls and mutation-style checks for selected repairs;
• documentation-currency checks derived from current source definitions.

This is a personal engineering project, not a claim of a customer production deployment or a multi-tenant managed service.

---

Technology stack

Area	Technology
Backend	Python, FastAPI
Frontend	React, TypeScript, Sigma.js, Vite, Tailwind CSS
Parsing	Tree-sitter plus custom domain extractors
Graph	KuzuDB and NetworkX
Vector retrieval	LanceDB
Lexical retrieval	BM25
AI integration	Claude, OpenAI-compatible providers, Gemini, Ollama and local HTTP inference paths
Agent integration	FastMCP
Testing	pytest and frontend Vitest suites in the private source repository

---

What an external reviewer can verify here

• the public architecture and domain model;
• the product flow from extraction to retrieval and UI;
• screenshots from the graph, Cypher and dashboard surfaces;
• interactive public demonstrations;
• the current source commit pin and date-bound capability inventory;
• explicit limitations and non-claims;
• preservation of the older showcase description for provenance.

The private source repository is not required to understand the problem, architecture or demonstrated user surfaces. It is required for a full source-code audit.

---

Documentation and demos

• Current evidence record
• Architecture deep dive
• Technical overview
• Architectural decisions
• Project context
• Search strategy demo
• Archived pre-refresh README

License

MIT