Skip to content
View soheilsayahvarg's full-sized avatar

Block or report soheilsayahvarg

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
soheilsayahvarg/README.md

Soheil Sayah Varg

Computer Engineering undergraduate at Sharif University of Technology, Tehran.

I work on inference and approximation under uncertainty — sampling from distributions that cannot be computed directly, and estimation for decision-making under partial observability.


Research

The confounded-POMDP implementation and results are public. The language-model manuscript is in preparation.

Exact controlled generation from tilted autoregressive targets Digital Media Lab, Sharif. Sampling from a reward-tilted language model exactly in distribution rather than approximately, via a particle-Gibbs ladder over a batched proposal chain. Supervised by Dr. Ali Rostami, in Prof. Hamid R. Rabiee's group. Manuscript in preparation.
Model-based RL in confounded POMDPs A public empirical implementation of a theory-only ICML 2024 result: bridge-function identification from negative-control proxies, oracle-verified to machine precision, with continuous-state experiments and documented negative results. I led the implementation and experimental evaluation in a three-person team.

Public work

Coursework and personal projects, cleaned up and documented.

Repository What it is
modern-information-retrieval Three phases: a search engine built from scratch (VSM/BM25/LM), embeddings and BERT fine-tuning, then a hybrid dense+sparse multimodal system with cross-encoder reranking.
convex-optimization-algorithms Chambolle-Pock and ADMM written from their update rules, checked against a solver rather than delegated to one.
market-predictability-study A profitable backtest, and the study of whether any of its stated reasons hold up. They do not.
segmentation-and-deep-rl Attention U-Net variants for segmentation, and Soft Actor-Critic built from scratch.
neogit A version control system in C. Staging, commits, branches, merge, tags, pre-commit hooks. No libraries.
atomic-bomber A JavaFX arcade game, built solo in a week during first year.

For other Sharif students

These two are not portfolio pieces. They are the full set of reports, schematics and simulations for two lab courses, published because I could not find anything to check my own work against when I took them.

Use them to check your work, not to replace it. The lab staff have seen these files.


Also: co-founder of Resonance, an online Physics Olympiad school for students who have no selective high school near them, and a teaching assistant at Sharif with 19 course-semester appointments, including appointments for Fall 2026.

📫 soheilsayahvarg@gmail.com

Pinned Loading

  1. model-based-rl-confounded-pomdps model-based-rl-confounded-pomdps Public

    End-to-end empirical implementation of model-based off-policy evaluation and pessimistic policy selection for confounded POMDPs (Hong, Qi & Xu, ICML 2024)

    Python 1

  2. war-of-attrition-rl war-of-attrition-rl Public

    RL agent for a multi-stage war of attrition with incomplete information. Bayesian type filtering + potential-based shaping. 1st among students on the class leaderboard.

    Python 1

  3. modern-information-retrieval modern-information-retrieval Public

    Three-phase information retrieval coursework: a from-scratch Goodreads search engine (VSM/BM25/LM), word embeddings and BERT fine-tuning, and a hybrid dense+sparse multimodal product search system …

    Jupyter Notebook 1

  4. market-predictability-study market-predictability-study Public

    A backtest returned several hundred percent. This is the study of whether any of its stated reasons hold up - martingale tests, OU estimation, multiple-comparisons calibration, and a randomised-ent…

    Jupyter Notebook 1

  5. convex-optimization-algorithms convex-optimization-algorithms Public

    First-order convex optimization implemented from the update rules and validated against a solver: Chambolle-Pock for TV denoising, ADMM for inpainting, basis pursuit with a phase-transition study, …

    Jupyter Notebook 1

  6. segmentation-and-deep-rl segmentation-and-deep-rl Public

    Attention U-Net variants for road segmentation and Soft Actor-Critic for continuous control, both built from scratch in PyTorch.

    Jupyter Notebook 1