Syed Abdul Basit

M.Sc. AI Engineering (thesis track), Okan University, Istanbul
Senior Machine Learning Engineer, Applied AI

When an AI system keeps learning after it is deployed, how do we check that it is still right?

I study LLM agents and multi-agent systems, and in particular how to verify them. My M.Sc. thesis uses LLMs as verification agents for systems that keep learning after deployment. At work I build the same kind of checks for production LLM pipelines.

Applying to PhD programs in the United States for Fall 2027.

The setting my thesis studies A deployed system learns from new data and proposes an update. An LLM verification agent checks the update. Accepted updates are deployed. Rejected updates go back to the learner with the reason. Deployed system learns from new data update LLM verification agent accept Ship update keep monitoring reject, with the reason The setting my thesis studies Deployed system learns from new data update LLM verification agent accept Ship update keep monitoring reject, with the reason
The setting my thesis studies.

Research interests

LLM agents and multi-agent systems: how agents are orchestrated, how they coordinate, and how they use tools.

Verifying and evaluating agents. Using LLMs as judges and verifiers, deciding when their verdicts can be trusted, and catching regressions before users do.

Agents that keep learning after deployment, and what stops them from quietly getting worse.

Healthcare as an application area. I spent four and a half years building ML and data systems in a hospital, and clinical settings are where unverified model updates cost the most.

M.Sc. thesis

Using LLMs as verification agents for continuous learning systems

Okan University, M.Sc. AI Engineering. Supervisor: Sina Alp. In progress.

The thesis looks at LLM agents that learn after deployment and uses LLMs as verification agents that check what those systems learn.

Publications

Bidirectional Cross-Modal Attention for Multimodal Disease Prediction

IEEE ASYU 2026, Istanbul. Accepted for oral presentation. First author. Supervised by Sina Alp.

Predicts disease from free-text symptom descriptions and vital signs together, with attention running in both directions between the two inputs. Experiments use synthetic patient data.

Code

PhyS2S-GNN

IEEE ASYU 2026, Istanbul. Accepted for oral presentation. First author, with Ahmed Anas. Supervised by Mahsa Mikaeili.

A physics-informed graph neural network for predicting compound extreme weather events. Experiments use synthetic data.

Research code

Token-efficient self-improving coding agent

LangGraph, local Ollama models, PostgreSQL. Work in progress.

A coding agent that retries its own solution, and an LLM judge that decides when another attempt is no longer worth the tokens. The question: can a cheap judge stop the loop early while keeping most of the pass rate of a full loop? The open problem is checking whether the judge's stop decisions are right. Results are not published yet.

Agent Memory OS

ChromaDB, Groq or Ollama, Streamlit.

A support agent with four memory layers: short-term, long-term, episodic and reflection. An experiment in what an agent should carry from one session to the next. No systematic evaluation yet.

Multimodal disease prediction

PyTorch, BioClinicalBERT.

Code behind the cross-modal attention paper: the model, three baselines, Monte Carlo dropout for uncertainty, and attention and SHAP explanations. Trained and tested on synthetic data.

Experience

Senior Machine Learning Engineer, Applied AI

Istanbul, March 2025 to present

  • Built a LangGraph multi-agent pipeline (Extractor, Validator, Router) that processes more than 10,000 documents a month.
  • Built a monitoring framework that runs RAGAS and LLM-as-judge checks in CI/CD.
  • Used LLM feature extraction to feed an XGBoost fraud-pattern classifier.

Machine Learning and Data Engineer, Aga Khan University Hospital

Karachi, September 2020 to March 2025

  • Built Spark and Azure data pipelines.
  • Built a clinical triage classifier fine-tuned with LoRA.
  • Managed a team of four.

Code written for employers belongs to them and is not public.

Open source

PyHealth pull request #1324

sunlabuiuc/PyHealth. Open, under review.

Fixes the documentation and shape check of regression_metrics_fn, returns plain Python floats, and adds 12 tests, including one that runs Trainer.evaluate on a regression model.

Education

M.Sc. AI Engineering (thesis track)

Okan University, Istanbul. 2025 to present. Taught in English.

B.E. Mechatronics

SZABIST, Karachi. 2014 to 2020.

Certifications: Claude Certified Architect – Foundations; AI Fluency (Anthropic Academy).

Contact

If you supervise PhD students in LLM agents, agent evaluation or clinical AI and are recruiting for Fall 2027, I would be glad to hear from you: basit121987@outlook.com.

Based in Istanbul. Languages: English, German, Urdu, and conversational Turkish.