About me
Hi! I'm Sogand Tatlari, an independent AI researcher and software engineer based in Tehran. I hold a B.Sc. in Computer Engineering from Hakim Sabzevari University.
My research interest is in why training and prompting fail in systematic ways, and how to design around those failures. So far this has covered two areas: optimizer design for neural networks on tabular data, and a controlled study of how prompt strategy interacts with the choice of LLM. I like experiments with careful statistical design, and I try to report limitations plainly.
Before research, I worked as a mobile developer and front-end team lead. Deploying machine learning models in production, such as a computer vision pipeline for ID verification, raised reliability questions that moved me toward research.
I am applying to thesis-based MSc programs for Fall 2027. I am currently studying reinforcement learning from Sutton and Barto and plan to test my optimizer in a small RL setting.
Selected work
LPECO, a control-inspired optimizer for tabular neural networks, and Emergent Brittleness, a factorial study of prompt strategy and model choice in LLM code optimization. Both are preprints with public code.
Publications
Unlocking Potential: A Case for Neural Network Perseverance on Critical Tabular Tasks Through Principled Optimizer Design
We introduce LPECO, which adds a PI-control-inspired integral correction to the Lion optimizer to damp excess momentum before the sign step. A ResNet-like network trained with LPECO was compared with Adam, AdamW, Lion, LightGBM, and XGBoost on eight tabular datasets under an equal 128-trial hyperparameter search. It was statistically equivalent to LightGBM (TOST) and had a 98.4% posterior probability of practical equivalence to XGBoost (Bayesian ROPE). On BreastCancer, its results were higher than both tree models (Mann-Whitney U, p < 0.001).
Limitations: eight datasets, one architecture, one random seed, accuracy only. The omnibus Friedman test was not significant (p = 0.422), the equivalence margin is derived from our own benchmark, and the correction term is a heuristic without a convergence analysis. The link between LPECO's advantage and class balance was a non-significant trend (ρ ≈ 0.54, p = 0.170).
Emergent Brittleness: A Case Study on Pathological Instruction-Following in LLM-Based Code Optimization
A 3×3×5 factorial experiment (270 runs) with three LLMs, three prompt strategies, and five synthetic Python optimization tasks, analyzed with a mixed-effects logistic model for correctness and generalized least squares for performance. An elaborate "expert" prompt significantly reduced correctness for gemini-2.5-flash-lite (p = 0.030) and showed a non-significant trend for qwen-2.5-72b-instruct (p = 0.098). A persona prompt showed a trend toward slower solutions for qwen-2.5-72b-instruct (p = 0.059). The effect of a prompt depended on the model.
Limitations: the expert prompt bundles several components, so I cannot say which one matters; tasks were synthetic; several effects are trends. Next steps are component ablations, more runs, and public benchmarks.
CV
Education
- Full scholarship.
- Capstone (20/20): domain-specific semantic search engine.
Research experience
- Two co-authored preprints (2025): LPECO (optimizer design) and Emergent Brittleness (LLM reliability), supervised by Prof. H. R. Hamidi (IKIU) for the second.
Professional experience
- Led a front-end team and built the company's React Native booking app.
- Designed and deployed two production ML models: gender verification from ID photos (computer vision) and offensive-comment detection (NLP).
- Built and maintained a React Native app for real estate listings.
Skills
- Machine learning: Python, PyTorch, TensorFlow/Keras, scikit-learn, Optuna (TPE)
- Statistics: mixed-effects models, GLS, TOST and Bayesian equivalence testing, Friedman and Mann-Whitney tests
- Engineering: React Native, React, JavaScript, REST APIs
Certifications
- Natural Language Processing; Custom and Distributed Training with TensorFlow (DeepLearning.AI, Coursera)