ML engineer / researcherDubai, UAE

Kasra Sartaee

Models that
know when
they’re wrong.

Evaluation and optimisation of LLM and generative AI systems.

View research Get in touch

Fig. 00 / From noise to structureIllustrative

Education

MSc AI & Data EngineeringUCL / 2026

BSc (Hons) Computer Science (AI)Heriot-Watt / First Class

01 / Research

Questions worth
getting right.

Evaluation, uncertainty, and the choices behind reliable AI.

Selected work / 2024–2026

01 UCL · MSc dissertation / 2025–2026

Multi-objective search over an LLM repair pipeline

MSc dissertation. Tuning the decoding and retry policy of a four-stage multi-agent LLM pipeline for automated program repair under three competing objectives: correctness, complexity and energy.

Random sampling beat NSGA-II on held-out correctness (Cliff's δ = 1, Holm-corrected p = 0.024).
Selecting on training correctness raised held-out correctness by 3.4–5.2 points in every arm–benchmark cell.
Correctness rankings transferred from C to Python (ρ = 0.63 over 665 policies); energy rankings did not.

Explore the research Source code

Python / Ollama · qwen2.5-coder 1.5B / pymoo / BoTorch / GPyTorch / Intel RAPL

NSGA-II
MOBO (qLogNEHVI)
Random search
dominated region, pooled front

Data: 2-D fronts

Non-dominated policies per arm, test-pass fraction and energy
arm	test-pass	energy (mJ)
NSGA-II	0.389	15.31
NSGA-II	0.392	15.34
NSGA-II	0.414	15.64
NSGA-II	0.502	15.66
NSGA-II	0.519	16.25
NSGA-II	0.543	16.38
NSGA-II	0.567	16.58
NSGA-II	0.571	16.72
NSGA-II	0.606	16.86
NSGA-II	0.616	18.00
NSGA-II	0.638	18.43
MOBO (qLogNEHVI)	0.274	14.96
MOBO (qLogNEHVI)	0.496	15.45
MOBO (qLogNEHVI)	0.575	16.09
MOBO (qLogNEHVI)	0.607	17.15
MOBO (qLogNEHVI)	0.616	17.57
MOBO (qLogNEHVI)	0.618	17.61
MOBO (qLogNEHVI)	0.619	17.73
MOBO (qLogNEHVI)	0.638	18.11
MOBO (qLogNEHVI)	0.639	19.10
Random search	0.379	15.43
Random search	0.504	15.65
Random search	0.521	15.96
Random search	0.538	16.19
Random search	0.579	16.23
Random search	0.611	17.26
Random search	0.622	17.64
Random search	0.626	18.33
Random search	0.636	18.48

4 seeds per arm (reps 1–4, one energy epoch) · 545 policies · 9 beyond 22 mJ not shown · reference point (22 mJ, 0.1) · source @ 3ad0d48

fig.1 Training Pareto fronts on Codeflaws (C), three search arms, 4 seeds eachreal data

02 UCL / 2025–2026

Probabilistic forecasting of geomagnetic indices

Transformer and LSTM forecasters of the Dst index with calibrated probabilistic outputs, benchmarked against a physics-informed baseline.

Probabilistic output heads, so each forecast is a predictive distribution rather than a point estimate.
Predicted uncertainty calibrated with temperature scaling.
Compared against a physics-informed hybrid baseline.

Explore the research

PyTorch

fig.2 Forecast fan chart with the median and 50% / 90% predictive bandssimulated / illustrative

03 Heriot-Watt · BSc dissertation / 2024–2025

EchoScript: script-aware music generation

BSc dissertation. Fine-tuning MusicGen to score film scenes from their scripts. Graded First Class; published in the AAAI Symposium Series.

Lower Fréchet Audio Distance than the base model (4.37 vs 4.55).
Listeners preferred it for both musical quality and narrative alignment.

Explore the research

PyTorch / Hugging Face / MusicGen

fig.3 Fréchet Audio Distance against the base model (lower is better)real data

Engineering

From ideas
to systems.

UCL / 2025–2026

Distributed ML inference platform

A 5-node Kubernetes inference cluster provisioned as infrastructure-as-code.

Kubernetes (K3s) / Terraform / Ansible / Docker / MinIO

UCL / 2025

Formal verification of Python standard-library code

Verifying CPython standard-library functions by symbolic execution, and testing where the method breaks.

CrossHair / Z3

02 / Publications

Written & published.

2 papers

AAAI-SS / 2025conference

EchoScript: Enhancing AI Music Generation for Cinematic Scoring via Script-Aware Fine-Tuning

Mohammad Kasra Sartaee, Kayvan Karim

Fine-tuned MusicGen for script-conditioned cinematic soundtrack generation; evaluated with Fréchet Audio Distance and structured listening tests.

Proceedings of the AAAI Symposium Series, 6(1), 323–332

BibTeX

@inproceedings{sartaee2025echoscript,
  author    = {Sartaee, Mohammad Kasra and Karim, Kayvan},
  title     = {EchoScript: Enhancing AI Music Generation for Cinematic Scoring via Script-Aware Fine-Tuning},
  booktitle = {Proceedings of the AAAI Symposium Series},
  volume    = {6},
  number    = {1},
  pages     = {323--332},
  year      = {2025},
  month     = aug,
  publisher = {AAAI Press},
  doi       = {10.1609/aaaiss.v6i1.36069}
}

IJDSAA / 2025journal

Uncertainty, Disagreement, and Information Fusion in Modern AI Systems: A Comprehensive Survey

Mohammad Kasra Sartaee, M. Q. Al-Shihani

A survey of uncertainty, disagreement and information fusion in AI systems.

International Journal of Data Science and Advanced Analytics, 7(2), 473–484

BibTeX

@article{sartaee2025uncertainty,
  author  = {Sartaee, Mohammad Kasra and Al-Shihani, M. Q.},
  title   = {Uncertainty, Disagreement, and Information Fusion in Modern AI Systems: A Comprehensive Survey},
  journal = {International Journal of Data Science and Advanced Analytics},
  volume  = {7},
  number  = {2},
  pages   = {473--484},
  year    = {2025}
}

03 / About

A little context.

I work on how to evaluate and configure LLM and generative AI systems reliably when their behaviour is stochastic and their objectives conflict: how to design and analyse experiments on randomised systems, how to search and select among configurations with competing objectives, how to quantify uncertainty and calibration, and what AI actually costs in compute and energy.

I have just completed an MSc in Artificial Intelligence and Data Engineering at UCL. My dissertation tuned the decoding and retry policy of a multi-agent LLM repair pipeline. Plain random sampling beat NSGA-II on held-out correctness, and how a configuration is selected from a Pareto front mattered more than which optimiser produced it.

Before that, I completed a First Class BSc in Computer Science (AI) at Heriot-Watt University, Dubai. My dissertation, EchoScript, fine-tuned a generative music model to follow film scripts and evaluated it against both objective metrics and listeners. I am also first author of a survey on uncertainty, disagreement and information fusion in AI systems.

Research interests

01Experimental design and statistics for stochastic AI systems
02Multi-objective optimisation and model selection for LLM pipelines
03Uncertainty quantification, calibration and evaluator disagreement
04Efficiency and energy measurement of AI systems

Recognition

2021–2025Deputy Principal's Award for Academic Excellence

Heriot-Watt University, four consecutive years

2025First Class Honours, BSc (Hons) Computer Science (AI)

Heriot-Watt University

20242nd place, Packaging Sustainability Symposium

Team lead, in partnership with Dubai Humanitarian

Teaching

Teaching Assistant, Functional Programming and Compiler Design

Heriot-Watt University, Dubai · 2024–2025

Tutorials and labs for 30+ undergraduates on lambda calculus, type systems and parsing, plus problem sets and marking rubrics. My first practical experience of evaluation design.

04 / What’s next

Let’s ask better
questions.

kasrasartaee@yahoo.com

I am looking for a funded PhD position on principled evaluation of AI systems, and I am also open to AI and research engineering roles. My CV has the details; feel free to get in touch.

kasrasartaee@yahoo.com github.com/Kasra-Sartaee LinkedIn Google Scholar ORCID CV

Kasra Sartaee

Models thatknow whenthey’re wrong.