← All work 01 · Personal project · Knowledge & agentic AI

Îktomnî knowledge engine

Raw documents in; a queryable knowledge graph, cited answers, and autonomous research reports out. A Go engine that fuses classical IR, knowledge-graph extraction, and an agentic reasoning loop — with epistemic grounding so the system knows what it actually knows.

Role
Directed the build with AI
Stack
Go · Neo4j · Ollama pool · PDFium
Domain
Agentic AI · RAG · KG
Status
Deployed on GPU box
0
Graph chunks
0
Concepts
0
Questions indexed
0
Reasoning stack
How I built this

I directed AI coding agents through the whole thing — architecture, implementation, and evaluation of a fairly deep system (agent loops, retrieval, graph reasoning). My job was framing the problem, making the design calls, and verifying the output was sound. The takeaway isn’t that I’m a Go expert — it’s that I can drive AI to produce a working knowledge engine and know enough to keep it honest.

Open the real Îktomnî Library

○ demo mode

This isn't a mock — it's the actual application I built, running live on my GPU box. Search across every book, ask cited questions, run the deep-research agent, and browse the knowledge graph. (Answers run on a home GPU, so give them a moment.)

Launch the live app
Interactive · demo 01

Watch an agent research itself to a stop.

The deep-research agent doesn't run a fixed number of steps. It runs a plan → retrieve → reflect loop and self-terminates on a composite stop condition: coverage and confidence met, or consecutive dry rounds, or a hard cap. Hit run and watch a live trace.

research-agent · plan-act-reflect trace ○ demo mode
idle
coverage0%
confidence0%

Every finding is conformal-gated before it's admitted — the agent abstains rather than assert when the evidence is a bare inferred edge. Grounding is a first-class layer, not an afterthought.

Interactive · demo 02

From a paragraph to a graph.

The extraction pipeline reads text, resolves entities, and links concepts with grounded, provenance-tracked edges. Pick a sentence and watch entities and relations light up.

extraction · entity + relation graph
Under the hood

A layered reasoning stack.

The pipeline

# ingest → understand → ground → reason PDFium render OCR / layout / tables extract entities + relations (LLM pool) resolve + canonicalize → Neo4j ground claims → epistemic tournament retrieve hybrid (BM25 + vector, RRF) reason → cited RAG / research agent

Engineering highlights

  • LLM pool — load-balances Ollama across a 2070 GPU box + M4 Mac Mini so heavy jobs run where the silicon is.
  • CPU/GPU pipeline — overlaps CPU rasterization with GPU recognition; race-tested, serial-output-equivalent.
  • Epistemic tournament — competing axiom-sets ranked above a sound deductive core; the system reasons about which beliefs to trust.
  • Layered QA — faithfulness, grounding, and stability gates in CI on bootstrap-CI lower bounds.
← All work