Portrait of Ilkin Isler

AI engineer and researcher building AI systems people can rely on.

  • Forward-deployed engineering
  • Trustworthy AI
  • PhD, Computer Science
Ask me anything Grounded answers

Hi! Ask me about my work, research, or background.

Answers come only from my resume, a UCF News feature, and first-party facts, and it says so when it doesn’t know.

More about me
About

Senior AI Engineer at Universal Creative

I work in a forward-deployed role: I meet with business teams, scope what they need, and take the AI solution from first prototype to trusted production. I lead an AI engineering team across our AI platform, finance, and cost-estimating work. Before that, my PhD focused on designing neural networks for medical image segmentation that know when they might be wrong: uncertainty estimates and explainability that both improved the model's accuracy and told clinicians when to rely on its output. That same question, when an AI answer can be trusted, drives the evaluation work I do today.

Role
Senior AI Engineer, Universal Creative
Education
PhD & MS in Computer Science, UCF
Focus
LLM agents & evaluation, RAG, medical imaging
Based in
Orlando, FL
Off the clock
Competitive powerlifter
Main initiatives

What I build at work

Most of my time goes to two forward-deployed initiatives: I embed with the teams, learn how they really work, and build the AI tools with them.

Universal Creative · Finance · Built from scratch

AI for finance reporting and forecasting

Finance teams kept a separate, frequently reworked spreadsheet for every project, so monthly reporting and forecasting meant hours of manual consolidation before any analysis could start.

What I designed

  • Human-in-the-loop AI tools, designed from scratch and built with my developers, that bring project finance data into one adaptable source of truth.
  • Monthly insights that surface the top cost drivers and what changed since last month.
  • Risk and contingency analysis that flags transactions worth a closer look.
  • Forecasting support that keeps the analyst in control of every number.

How I approached it

As solution architect, I started from how analysts actually close a month, not from what the model could do, and led the developers who built the tools. Every output stays reviewable and editable, and new methods go through evaluation before they reach the team.

What changed

Cash Flow & Earned Value is in production with 70+ finance users globally, saving 15–30 minutes per analysis and about an hour per forecast. Monthly Update Insights and Risk & Contingency have their first versions in production, with more capabilities in development.

Human-in-the-loopLLM toolsForecastingEvaluation
Universal Creative · Estimating · Built from scratch

AI-assisted cost estimating

Estimating large, one-of-a-kind projects depends on years of expertise and messy, non-standardized inputs. The estimators had seen ambitious technology efforts come and go, and they were rightly skeptical.

What I designed

  • The Estimator Builder: a suite of 12 purpose-built, human-in-the-loop AI applications for ~80 users, designed from scratch, with 6+ already shipped.
  • Drawing Intelligence: reads design drawings to draft project scope.
  • Concept Estimate Automation: turns messy, non-standardized scope lists into structured estimates and tracks how they change over time.
  • Pricing Intelligence: retrieves everything previously paid for any item across more than 2 million files.

How I approached it

I lead the initiative and its developers. Being an expert in technology doesn't make you an expert in someone else's domain, so we replicated existing workflows first: the tools accept the estimators' own Excel files and return results in the same familiar format. We started with small wins they validated, added automation as trust grew, and built shared infrastructure underneath so the step-by-step approach never capped what we could deliver.

What changed

The conversation moved from whether it would work to what else we could automate, and estimating that used to take days now takes minutes, and it's traceable and consistent.

Forward-deployedDocument AIRetrievalAdoption
In progress · Built from scratch

Project intelligence: a temporal knowledge graph

Leadership needs to know what changed on a project and what it affects, but the answer is spread across design, scheduling, estimating, budgeting, and finance systems and a steady stream of project decks.

What I'm building

A project-intelligence tool built from scratch on a temporal knowledge-graph architecture similar to Graphiti:

  • Every new deck, schedule, or budget update is ingested as an episode, so each fact traces back to its source.
  • An LLM extracts project entities and the relationships between them into a graph.
  • Facts are never overwritten: when something changes, the old version is closed out with its validity window, so the system knows both what is true now and what was true at any point.
  • Retrieval combines semantic search, keyword search, and graph traversal, so "what changed this month, and what does it affect?" follows the links from a scope change to the budgets and schedules downstream.

The goal

One view for senior leadership that shows what changed across a project and how it affects budgets, plans, and schedules, with every answer traceable to the document it came from.

Knowledge graphsTemporal dataHybrid retrievalProvenance
Work

Selected work

Universal Creative

Enterprise AI systems

An in-house platform for AI agents and multi-agent workflows, LLM evaluation and hallucination detection pipelines, and human-in-the-loop tools that help business teams move from spreadsheets to AI workflows they can trust.

PhD research · UCF

Uncertainty-aware medical image segmentation

Novel 3D architectures and loss functions for tumor and organ-at-risk segmentation, with uncertainty estimates that tell clinicians when to rely on the model. Raised Swin UNETR Dice from 0.47 to 0.64.

See the code
This site

An assistant that knows its limits

The chat above answers only from my resume, a UCF News feature, and first-party facts. It uses hybrid keyword + embedding retrieval, declines instead of guessing when it isn't confident, and checks every sentence against its sources.

See how it works
Research

Publications

Full list on Google Scholar

Get in touch

Happy to talk about trustworthy AI, LLM evaluation, or research collaborations.

Email me Resume LinkedIn