Samira Korani Knowledge Graphs · GraphRAG · LLM Evaluation

LLM evaluation and knowledge graph consulting

Find out where your AI system fails and what to fix.

I help engineering teams evaluate RAG and agent systems, investigate unreliable outputs, and make better decisions about knowledge graphs and production AI.

For startups and mid-sized businesses anywhere in the world that are building AI features, especially in fintech and insurtech, and for teams already using Neo4j. I work remotely from Galway, Ireland.

Book a call See the packages

Featured package

LLM Evaluation Health Check

Price
EUR 2,950 excl. VAT Fixed project fee
Time
About 5 working days After access and scope are agreed
Scope
One RAG or agent workflow, up to 50 test cases
What's included

Public work you can check

Services

Fixed-price packages for LLM evaluation and knowledge graphs

I do the work myself, so you deal with one person from the first call to the handover.

Featured package

LLM Evaluation Health Check

Find out where one RAG or agent workflow fails, why, and what to fix first.

Who it's for

Engineering teams with a RAG or agent feature in production, or close to release, who are seeing answers they can't fully trust, or who need evidence before a model change, a launch or a customer's due-diligence questions.

What you get

  • Agreed evaluation cases (up to 50), written with your team
  • Baseline results across the agreed measures
  • A failure analysis showing where and why the workflow fails
  • Prioritised fixes, ordered by likely impact and effort
  • A 60-minute handover with your team
Price
EUR 2,950 excl. VAT Fixed project fee
Time
About 5 working days After access and scope are agreed
Scope
  • One existing RAG or agent workflow
  • One agreed evaluation environment
  • Up to 50 representative test cases
Book a call Email me

Full package details

Other packages

Engagements are contracted with and invoiced by MILEAWAY AI SYSTEMS LIMITED. Prices are in euro and exclude VAT, which is added where applicable. Any third-party costs, such as model or cloud usage, are confirmed in the quotation. All packages

Process

How an engagement works

  1. Discovery

    A short call about your system, what's going wrong and what a useful result would look like. If the work isn't a good fit, I'll say so.

  2. Agreed scope

    A written scope and fixed-fee quotation covering the workflow, the environment, the test cases, access and timing.

  3. Assessment

    Once access is in place, I build the evaluation cases with your team, run the baseline and analyse the failures.

  4. Handover

    A 60-minute session to walk through the results and priorities. You keep the evaluation cases and the report.

Work

Selected technical work

Open-source projects, talks and write-ups. These are my own demonstrations and research, not client case studies.

Talk and write-up

Contextual knowledge graphs for grounding and evaluating LLMs

A knowledge graph that combines Geonames, OpenStreetMap, Wikidata and venue data to give LLMs real-world context about places, used both to ground a travel assistant's answers and as ground truth for evaluating a fine-tuned model.

  • Neo4j
  • Knowledge graphs
  • LLM evaluation

Open-source prototype

Insurance policy RAG prototype

A retrieval-augmented assistant over insurance policy and regulatory documents. Documents and chunks are modelled in Neo4j alongside a native vector index, so retrieval can combine similarity search with category filters in Cypher. A local model served through Ollama generates answers, behind a FastAPI service and a Streamlit interface.

Status: Work in progress. Several modules, including evaluation and tests, are still placeholders.

  • Python
  • Neo4j
  • RAG
  • FastAPI

Hackathon prototype, Ecohack 2025

EcoSci, an LLM agent over a research knowledge graph

Pulls an ecology topic's links to publications from a Neo4j graph, serialises the subgraph for an LLM and asks it to suggest areas for further investigation, with a small Streamlit front end.

  • Python
  • Neo4j
  • Agents

Earlier open source

Persian NLP tools and open datasets

A semantic tokenizer for normalised Persian text, released under the MIT licence, and public datasets for students and researchers covering Persian tweets, named entities and banking statistics.

  • NLP
  • Datasets

More on GitHub

Writing

Notes from building with graphs and LLMs

All writing and research

Speaking

Speaking and workshops

I speak about evaluating LLM systems and putting knowledge graphs to work, for conferences, meetups and engineering teams.

NODES 2025 (Neo4j) ·

Building Contextual Knowledge Graphs to Adapt, Evaluate, and Leverage LLMs

How I combined Geonames, OpenStreetMap, Yelp and Wikidata into contextual knowledge graphs that give LLMs real-world context, and used them to adapt and evaluate model output. The session was live-streamed from NODES 2025.

This Week in Neo4j ·

Neo4j Featured Community Member

Featured in Neo4j's weekly community round-up, which also highlighted my NODES 2025 session.

Topics

  • Finding where your RAG system fails Building evaluation cases from real usage, checking groundedness, and turning failures into a prioritised fix list.
  • Graph, vector or hybrid? A practical way to decide when GraphRAG earns its complexity, and when better chunking or metadata would do more.
  • Contextual knowledge graphs for grounding LLMs Combining open datasets into a graph that gives models real-world context, building on my NODES 2025 session.
  • Observability for LLM applications Tracing, latency and cost, and catching regressions before users do.

Workshops

  • LLM evaluation for engineering teams A hands-on session in Python on building a small evaluation set and harness for a sample RAG application.
  • Knowledge graphs and GraphRAG with Neo4j A practical introduction to modelling, loading and querying a graph for retrieval.

Workshops are available for teams and events on enquiry.

Invite me to speak Talk history
Samira Korani

About

I build the kind of systems I evaluate

I'm Samira Korani, a machine learning engineer based in Galway, Ireland. I've worked in natural language processing and machine learning since 2013. Today I focus on production AI: evaluating LLM applications, building knowledge graphs and GraphRAG systems, and designing agentic workflows in Python, with the observability to see what they are doing.

My experience includes AI work in financial services and insurance, and co-founding Tripoh, a travel assistant built on LLMs and a tourism knowledge graph. My research background is in relation extraction and knowledge graph construction, and I have taught on the Knowledge Graphs (2022) and Programming I and II (2023) modules at NUI Galway, now the University of Galway.

I grew up in Iran, where a long-running question, how to live my life my way, took me through languages, cultures, music and dance before it led me to technology.

At a glance

Focus
LLM evaluation, knowledge graphs, GraphRAG, agentic workflows
Sectors
Financial services, insurance, travel
Tools
Python, Neo4j and Cypher, FastAPI, Docker, Kubernetes, AWS, Google Cloud
Community
NODES 2025 speaker; Neo4j Featured Community Member
Based in
Galway, Ireland

My full work history is on LinkedIn, and my code is on GitHub.

LinkedIn profile GitHub profile

FAQ

Questions teams usually ask

What do you need from us for a Health Check?

Access to the workflow in one agreed environment (often staging), a set of real or realistic questions or tasks, any examples of bad outputs you have already seen, and a technical contact who can answer questions during the week.

Are 50 test cases enough?

Enough to find systematic failures and decide what to fix first, if the cases are chosen well. Not enough to make statistical claims about production behaviour, and I won't present the results as if they were. The cases are yours to extend afterwards.

Do you also fix what you find?

The Health Check ends with prioritised fixes and a handover. If you'd like help implementing them, I'll scope that separately, so you can decide with the results in hand.

Which stacks do you work with?

I work mainly in Python, with Neo4j for knowledge graphs. The Health Check evaluates your workflow as it runs, so it doesn't depend on a particular framework or model provider.

Do you work with teams outside Ireland?

Yes. I work remotely with startups and mid-sized businesses anywhere. Calendly shows call times in your own time zone, and prices are in euro.

How do you handle confidential data?

Access, data handling and confidentiality terms are agreed in writing as part of the scope, before any work starts. I only ask for the access the agreed work needs.

Who do we contract with?

You work directly with me. Engagements are contracted with and invoiced by MILEAWAY AI SYSTEMS LIMITED, the company through which I provide consulting services. Prices are in euro and exclude VAT, which is added where applicable. Any third-party costs, such as model or cloud usage, are confirmed in the quotation.

Contact

Tell me about your system

Book a 30-minute call, or send a few lines about what you're building and what's going wrong.

Book a call Email me about a project Speaking or workshop enquiry

Email