MSCS, Georgia Tech — machine learning and NLP for financial markets

Tunis, Tunisia

Karim Razgallah — PhD applicant in financial engineering

I work on pulling quantitative signal out of unstructured corporate disclosure, mostly SEC filings, using natural language processing and machine learning. Alongside that I work on derivative pricing under market frictions, and on whether LLM-based agents are dependable enough to be trusted in financial applications.

Applying — Fall 2027

EmailCV (PDF)

RESEARCH

[01] Selected work

01In progress

Option valuation under daily price limits

ProblemExchanges halt trading once a price moves past a daily limit, but standard option pricing models assume the underlying can move freely, so they misprice derivatives written on limit-constrained assets.

ApproachStatistical models of asset price volatility under exchange-imposed daily limits, together with pricing models adjusted for the constrained underlying. Pricing is done by Monte Carlo simulation and dynamic programming, and validated empirically on markets that actually use limit mechanisms.

derivative pricing
numerical methods

02In progress

Predicting mutual fund flows and performance

ProblemFund flows and alpha are normally predicted with linear factor models, which are the benchmark any alternative has to beat on data it has not seen.

ApproachA two-step hybrid of linear models and neural networks for predicting fund flow and alpha. Evaluated under walk-forward validation against both linear and factor-model benchmarks, so the comparison is made out of sample rather than on a single fit.

machine learning
asset management

03In progress

Structured corpora from SEC fund disclosure

ProblemMutual fund annual reports are filed as free text and their layout varies by fund and by year, so the disclosure cannot be used at scale until it has been parsed into something consistent.

ApproachAutomated extraction and parsing pipelines over SEC EDGAR that turn filings into structured corpora for NLP training. The dataset covers roughly 15,000 filings from about 1,000 funds across 17 years, and underpins an ongoing programme applying NLP to fund disclosure for performance evaluation.

nlp
asset management

04Under review

MAM: Multimodal Agentic Memory

ProblemLLM agents lose track of what came earlier in a long task, which surfaces as hallucination and as reasoning that contradicts itself.

ApproachA multimodal memory and retrieval framework for LLM-based autonomous agents, designed to reduce hallucination and hold reasoning consistent over long horizons.

llm agents
machine learning

05In progress

Binary patch presence testing

ProblemDeciding whether a security patch is present in a compiled binary when the source code is not available.

ApproachApplying large language models to the problem. The core difficulty is representation under distribution shift: identical semantics compiled under different toolchains, optimization levels and architectures produce very different surface forms.

llm agents
security

PUBLICATIONS

[02] Reverse chronological

[01]
Valuation of options and volatility estimation under daily price limit constraints
Karim Razgallah, Hatem Ben Ameur, E. Errais, M. Turki
Presented at the GIGS 4th Annual Conference, Global Initiative for Governance and Sustainability, 2025
Targeted at the Journal of DerivativesyearIn preparation
[02]
A hybrid model approach to predicting mutual fund performance
Karim Razgallah, Hatem Ben Ameur, S. Lazrak, M. Ayadi, A. Trabelsi
venueyearWorking paper
[03]
MAM: Multimodal Agentic Memory for LLM Agents
Karim Razgallah, V. K. Madisetti
ACM AI LettersyearUnder review

FOCUS

[03] Directions for a PhD

01
Quantitative signal in corporate disclosure
Funds and firms say a great deal in prose that never reaches the numeric datasets everyone trains on. The obstacle is not that the text is hard to read, it is that the same disclosure is laid out differently by every filer and re-laid-out again every few years, so any pipeline built against one format quietly rots. I want to work on extracting quantitative signal from filings at scale, and on evaluation that separates genuine predictive content from the effects of formatting and survivorship in the corpus itself.
02
Models that see through surface form
A filing reformatted and a function recompiled have the same property: the semantics are unchanged while the surface form is not. Models trained on one presentation degrade on another, and the usual evaluation hides this because training and test data are drawn from the same source. I want to study representations that survive that shift, and evaluation protocols where held-out means a held-out filer or a held-out toolchain, not a held-out sample.
03
Reliability of LLM agents in financial applications
An agent that is right most of the time is not usable in a regulated setting unless you can say which cases it is wrong in. Long-horizon tasks make this worse, because an agent that has lost track of earlier context produces confident output that contradicts what it established a step ago. I am interested in memory and retrieval that keep reasoning consistent, and in calibration good enough that an operator can set a threshold against a stated error budget.
04
Derivative pricing under market frictions
Standard pricing theory assumes an underlying that can move freely, and real exchanges routinely break that assumption: daily price limits halt trading exactly when the information arrives. The interesting question is what happens to volatility estimation and to the value of a contract when the observed price is censored by a market rule rather than by the economics. I want to keep working on pricing and numerical methods where the friction is part of the model instead of an afterthought.

EXPERIENCE

[04] Research, teaching, industry

Research

Graduate Researcher
Cyber-Physical Systems Security Lab, Georgia Tech
Applying large language models to binary patch presence testing: deciding whether a security patch is present in a compiled binary when source code is unavailable.
Jun 2026toPresent
Independent Research (Special Problems)
Georgia Institute of Technology
Graduate research course supervised by Prof. V. K. Madisetti. Designed and implemented a multimodal agentic memory framework for large language models, addressing long-horizon reasoning consistency and hallucination. Resulted in a working paper co-authored with the supervisor.
Jan 2026toMay 2026
Research Collaborator
HEC Montréal · Brock University · Université de Sherbrooke
Ongoing collaboration with Prof. Hatem Ben Ameur spanning NLP and machine learning for asset management. Building automated extraction and parsing pipelines over SEC EDGAR filings, producing structured corpora of fund disclosure for NLP model training.
Jun 2024toPresent
Research Assistant
HEC Montréal
Hand-collecting data from SEC filings to build a database of mutual fund annual reports, under Prof. Hatem Ben Ameur. Roughly 15,000 filings covering 1,000 funds over 17 years, underpinning an ongoing programme applying NLP to fund disclosure.
Sep 2023toDec 2023

Teaching

Teaching Assistant
Tunis Business School
Dynamic Asset Pricing Theory (graduate), appointed after completing the course as an undergraduate: tutorials on stochastic calculus, martingale pricing and continuous-time models. Numerical Methods in Finance (graduate): prepared materials and Python implementations for Prof. Hatem Ben Ameur during his visiting appointment.
Sep 2022toJun 2023

Industry

Software Engineer
Self-employed, freelance
End-to-end software for clients in the USA and Canada, across the full lifecycle including AWS deployment and CI/CD.
Jun 2024toPresent
Quantitative Analyst
Colombus Capital
Forecasting models for currency movements and macroeconomic factor analysis using Refinitiv Eikon. Integrated statistical and machine learning methods into FX trading models.
Jan 2024toJun 2024
Software Engineer
Expertify, Paris
Secure, scalable architectures with multi-role authentication and access control, and real-time dashboards over multi-source data.
Feb 2022toDec 2023

Education

M.S. Computer Science
Georgia Institute of Technology
Concentration in Artificial IntelligenceGPA 4.0
Jan 2026toPresent
Research stay
HEC Montréal
Department of Decision Sciences and GERAD
Financial engineering research under Prof. Hatem Ben Ameur. Option pricing and numerical methods, one semester in residence in Montréal. Basis of the undergraduate thesis.
Jun 2023toJul 2024
B.S. Business Administration
Tunis Business School
Major in Finance, minor in Information Technology
Thesis: “Valuation of European and American Options under Daily Price Limit Restrictions”. Supervised by Prof. E. Errais (TBS) and Prof. H. Ben Ameur (HEC Montréal, external).
Sep 2019toJun 2024

CONTACT

[05] No form

I am applying for PhD positions in financial engineering starting Fall 2027. If your group works on machine learning for asset management, empirical finance with unstructured data, or derivative pricing, I would like to talk. Email is the fastest way to reach me.

[email protected]CV (PDF)

Georgia Tech[email protected]

Option valuation under daily price limits

Predicting mutual fund flows and performance

Structured corpora from SEC fund disclosure

MAM: Multimodal Agentic Memory

Binary patch presence testing

Valuation of options and volatility estimation under daily price limit constraints

A hybrid model approach to predicting mutual fund performance

MAM: Multimodal Agentic Memory for LLM Agents

Quantitative signal in corporate disclosure

Models that see through surface form

Reliability of LLM agents in financial applications

Derivative pricing under market frictions

Research

Graduate Researcher

Independent Research (Special Problems)

Research Collaborator

Research Assistant

Teaching

Teaching Assistant

Industry

Software Engineer

Quantitative Analyst

Software Engineer

Education

M.S. Computer Science

Research stay

B.S. Business Administration