What I build
Retrieval pipelines
Chunking, hybrid search and rerankers tuned against your own documents, not a benchmark.
Knowledge graphs
Neo4j schemas that turn scattered files into entities and relations a model can traverse.
Agents in production
Tool use, guardrails and evaluation harnesses, deployed where your team already works.
Live demo
An assistant built on my own project history. Ask what I work with, or whether I have built something like your project — it answers from real work, and says when I haven't.
Do you have experience with graph-based retrieval?
Yes — Neo4j and pgvector, two production systems. Graph traversal for context, embeddings for recall.
About me
I build backend systems and the AI layers that sit inside them. ExcaliburSoft is one engineer, which means the person who designs the schema is the person who writes the migration, and the person you talk to when something goes wrong.
The work is usually a platform with real users and awkward data behind it — multi-tenant SaaS with per-tenant isolation, a B2B procurement marketplace, invoice reconciliation matched against ERP records. NestJS and PostgreSQL underneath, Neo4j and pgvector where relationships and recall actually matter.
I take two or three projects at a time, usually two to eight weeks each, and I plan the whole thing in detail before writing any code. If a retrieval layer or an agent is the wrong fix for your problem, I will say so — that conversation costs less than the build.