Dr. Bedir Yilmaz, a computer-vision PhD turned AI Architect at lcmd, explains how the company's first attempt at a construction-site conversational agent — a RAG pipeline built over AI-generated prose narratives of project activity — could answer "what already happened" questions but failed at the sense-making questions site experts actually needed, like what's threatening a project's timeline. That gap pushed him toward a Neo4j knowledge graph, built using agentic construction techniques, which now spans 108 projects with roughly 25 million nodes and 83 million relationships in private beta. The graph powers "Consultant in a Box," a natural-language agent that generates Cypher on the fly and surfaces fleet-wide construction wisdom — like which trade typically precedes drywall — without exposing any single customer's private project data.

Topics Covered

• lcmd: a Stuttgart-based lean construction management platform spun out of Drees & Sommer, used on complex multi-year builds like Germany's largest particle accelerator for antiproton research and the Denmark–Germany rail tunnel
• From computer vision PhD to construction AI: Bedir's path from crowd counting/density estimation research into an AI Architect role with no prior construction background
• The original goal: a conversational engine letting construction site teams talk to their own project data
• Why the app's collaborative, highly efficient source data wasn't usable by LLMs in its raw form, and the first fix: using coding agents to transform it into prose narrative corpora for a traditional RAG pipeline
• Where RAG hit its ceiling: it could answer "what already happened" questions (who's the most influential person, what happened last month) but site experts needed sense-making questions about the past and future — like what's threatening the project timeline — that similarity search can't answer because it ranks by linguistic closeness, not construction-domain logic
• The graph moment: realizing the missing connections were logical, not linguistic, and that a knowledge graph was the way to represent that construction logic
• Building the graph with agentic construction, inspired by Andreas Kollegger's Coursera lectures on agentic knowledge graph construction, and reaching near-parity with lcmd's own API within about an hour of applying what he'd learned
• Current scale: 108 projects loaded into one connected graph (~25M nodes, ~83M relationships) — in private beta, not yet GA
• Why a shared, multi-tenant graph works: project schemas are structurally similar (trades, tasks, dependencies), letting lcmd answer cross-project "wisdom of construction" questions without surfacing any customer's private data
• "Consultant in a Box" — lcmd's natural-language agent, built on a custom text-to-Cypher engine plus pre-built query templates for anticipated questions, with text-to-Cypher reserved for the ones it can't anticipate
• Why entity resolution hasn't been a problem: the knowledge graph is built deterministically from lcmd's own structured application data (trade selection, task dependencies, start/end dates), not extracted from unstructured text
• lcmd's product thesis: the knowledge graph is the product, not just a means — the long-term vision is letting other people's agents query lcmd's construction knowledge at a fair market price
• lcmd is now an official member of the Neo4j Startup Program and is already putting its Aura credits to use

Learn more about the ⁠Neo4j Startup Program

Alex Dollery is co-founder and Chief Technology Officer of Taskd, which is building connected company knowledge that brings facts, relationships, rules and evidence together for people and AI agents. His focus is making that knowledge a lasting business asset that remains useful as models, tools and workflows change. He first encountered Neo4j around twelve years ago at BenevolentAI, working on a biomedical knowledge graph for drug discovery. He later served as a director of AI and Machine Learning at the London Stock Exchange Group.

Taskd connects business information, expert rules and operational history so AI workflows can act reliably, explain their work and keep people in control of consequential decisions.
In this episode, Alex explains why relational databases break down once you're storing many different shapes of data and need AI to reason over how it all connects — and how Taskd wraps logical rules and "concept graphs" around raw facts in Neo4j to produce explainable, evidence-backed answers. He walks through a real government hackathon build that turned 14 siloed databases into a rules-driven graph to flag suspicious multi-company transactions, then demos Sage, Taskd's upcoming natural-language query engine that answers questions from a document graph without leaning on a large LLM. Taskd is now working with paying customers — including a maritime industry client at 97.3% accuracy — and is looking to talk to teams that need to trace AI-driven decisions back to verifiable source data.

• What Taskd does: merging a graph database with logic to provide an evidence trail behind facts, answers, and decisions
• Why relational databases (Postgres) break down once data gets more varied in shape, and why a graph gives the flexibility to model it
• Moving beyond ontology into "concept graphs" — layered concepts and the relationships between them, not just raw data
• Why graph traversal makes it possible to infer connections across nodes that would require painful joins in a relational database
• Live demo: the Alberta government hackathon — 14 siloed databases turned into a graph (~9M nodes, ~38M with derived concepts), layered with rules to flag suspicious inter-company transactions and multi-board-member ownership patterns
• Agents that continuously enhance and expand the concept graph over time
• Bitemporal modeling — reasoning about what was known "as of" a past point in time vs. what's known now, and why that matters for financial/quant use cases
• Live demo: Sage, Taskd's upcoming query engine — answering natural-language questions from a news article corpus by traversing the concept graph, without a large LLM, and tracing every answer back to source evidence
• The neurosymbolic approach: small fine-tuned language models paired with a structured graph, rather than large frontier LLMs or GraphRAG, for reliable and cost-efficient reasoning
• Support from Mila (the Quebec AI research institute) and the city of Montreal
• Where Taskd is now: paying customers (97.3% accuracy with a maritime industry client), actively growing, and looking to talk to teams that need explainable, source-traceable AI decisions without heavy LLM costs

Learn more about the Neo4j Startup Program

Dr. Yermie Cohen, MD, is the Founder and CEO of Memorang. Over 15 years in education technology, he has developed the curriculum and technology behind hundreds of test-prep programs reaching millions of learners across healthcare, professional licensure, and English proficiency. He pioneered machine-learning-powered spaced repetition in the early 2010s and launched one of the first applications of LLM-generated educational content in early 2022. Today, he builds multi-agent systems and Knowledge Graphs for high-stakes assessment and adaptive learning. He holds multiple engineering degrees from MIT and a Doctor of Medicine from UCLA.

Memorang is building **the AI stack for education**, a vertically integrated platform for adaptive learning and assessment. After winning the inaugural Vercel AI Accelerator from a field of thousands of companies, Memorang built an enterprise platform that powers high-stakes assessments and personalized learning experiences at scale. Its Neo4j-powered Knowledge Graph transforms fragmented curricula, content, assessments, and learner data into structured, queryable institutional memory. AI agents use that foundation to automate curriculum development, content generation, psychometric evaluation, and personalization, while web and mobile apps bring those capabilities to market. Memorang is backed by ETS Capital.

Yermie has been thinking in graphs since medical school — he'd sketched a graph model for pharmacology and anatomy before graph databases existed as a cloud service. When he set out to build Memorang's AI stack for education, he started with Neptune and moved to Neo4j for Cypher's standardization, deeper traversal performance, and managed hosting. The result: learner-data retrieval that took 8–9 seconds in Postgres now runs in under 500 milliseconds.
The deeper insight is how Memorang uses the graph not just to store curriculum and learner data, but to control every AI agent in the system. Rather than relying on semantic matching to route tools, sub-agents, and evaluators — which Yermie puts at around 70–80% accuracy — Memorang encodes subgraph membership in each skill's metadata so the right procedural memory and the right evaluator are called deterministically. For high-stakes assessment, where agent output has to be legally defensible in court, that precision isn't optional.

Topics Covered:

• Building Memorang's knowledge graph from a systems-engineering mindset — why relational databases break at depth two or three in curriculum modeling
• From Neptune to Neo4j: Gremlin's lack of standardization, missing managed hosting, and the hiring challenges that drove the switch
• The knowledge graph ontology: thick branches (subject areas), topical nodes, artifacts (items), and user-interaction edges
• Item response theory via nearest-neighbor graph traversal: predicting mastery on content a learner has never seen
• Infinite content generation + graph-weighted curriculum = ROI-based study scheduling
• Graph-controlled agent routing: encoding subgraph membership in skill YAML front matter to replace semantic tool calling with deterministic dispatch
• Reducing learner-data retrieval from 8–9 seconds to ~500 ms using Neo4j traversal optimization, caching, and query parallelization
• Hybrid architecture: Supabase (Postgres) + Neo4j + Redis, kept in sync via database triggers, each doing what it does best
• Applying conditionally selected evaluators to agent traces for high-stakes, legally defensible assessment content
• Hill climbing from 50% to 90% agent-output alignment with fewer than 100 human annotations by optimizing narrow, graph-scoped context rather than monolithic system prompts
• Roadmap: high-profile enterprise launches through end of 2026, general-availability developer platform in early 2027