mahdi afshari

-Coursework: ML, Computer Systems, Software Construction, Algorithms

• -Designing an RL loop that trains a policy to acquire and inject task-relevant knowledge into a frozen LLM's context (later LoRA weights), scored by forced-choice accuracy gains over fixed-prompt and cold baselines on held-out questions.
• -Built the ReST-EM training pipeline (candidate sampling → retention gating → SFT) with reward-reliability and train/eval leakage-safety gates required before any training run.
• -Sole researcher on the implementation, with weekly meetings with the PI and Omar Khattab observing as a potential future collaborator.

• -Built the Polymarket data pipeline (deduplication, side-flip filtering, resolution tracking) and a 35-configuration ablation harness spanning model, temperature, persona, and information-access axes, evaluated on 822 resolved binary questions.
• -Found the best single LLM forecaster earned significantly higher ROI than the human-crowd aggregate (+4.0% vs. −0.5%, p=0.004) at statistically indistinguishable accuracy (77.9% vs. 79.9%).
• -Showed that filtering to medium within-crowd agreement doubled ROI to +9.1%. Manuscript under review.

• -Built autonomous negotiation systems where LLM agents form alliances, trade resources, and coordinate without human mediation in Minecraft-like environments.
• -Extended MCrafter with 4 social primitives (alliance, trade, truce, protection) enabling 10+ concurrent agents with competing goals to negotiate autonomously.
• -Designed episodic memory and summarization layers to keep inference cost manageable while maintaining context across multi-turn negotiations.
• -Built evaluation infrastructure tracking diplomacy outcomes (alliance stability, resource distribution fairness) and construction progress in multi-village scenarios.

• -Designed and evaluated Drift Diffusion Models (DDMs) and their variants to capture behavioral patterns in short- and long-term memory tasks using data from 500+ trials per subject (human and non-human primates).
• -Integrated Gaussian Process (GP) regression into the modeling pipeline to estimate time-varying latent parameters, enabling non-parametric inference of cognitive dynamics and uncertainty quantification across trials.
• -Found that encoding dynamics degrade predictably with delay, but recall exhibits non-monotonic threshold shifts suggesting distinct neural mechanisms.
• -Applied probabilistic modeling techniques to analyze response distributions, jointly fitting reaction times and accuracy under varying memory loads and delays. Implemented models in JAX for GPU-accelerated parameter estimation.

• -Developed a novel approach to predict protein function from secondary structure as part of a team.
• -Fine-tuned ESMFold embeddings for secondary structure inputs and implemented an iterative feedback loop that refines structure predictions by incorporating function prediction errors, creating a bidirectional refinement cycle between structure and function models.
• -Improved function annotation accuracy by 15% on benchmark set using PyTorch and HuggingFace Transformers.

• -Teaching digital logic design, Verilog, and Minispec (Bluespec derivative) to 200+ students in weekly labs and office hours.
• -Debug student processors in real-time-pipelined CPUs, cache hierarchies, hazard detection-with average fix time under 10 minutes.
• -Release labs, write quizzes, and maintain course infrastructure (grading scripts, lab checkoff systems, debugging tools).
• -Help students understand not just how to write Verilog but why their designs work (or don't)-timing, synthesis, hardware intuition.

• -Led MIT's largest cultural umbrella organization, representing 1000+ international students across 15+ cultural groups.
• -Organized 4 large-scale events with 300–1100 attendees each ($50k+ total budget): cultural showcases, networking mixers, orientation programs.
• -Designed coordination framework for 15+ organizations-shared calendars, budget allocation system, conflict resolution processes-now used campus-wide.
• -Shipped new constitution establishing clear governance structure, term limits, and transition protocols for future leadership.
• -Built mentor network connecting 200+ incoming international students with upperclassmen, faculty, and startup founders for career guidance.

• -Built Siamese network for music similarity on 10k+ track dataset in a selected collaborative team of five.
• -Optimized triplet loss training pipeline, improving recommendation accuracy from 60% to 78% in blind A/B tests with 50 users.
• -Integrated Spotify API for metadata retrieval and React frontend for user interaction.

Gold Medal, ranked #7 globally among 300+ participants from 75+ countries in Yerevan, Armenia; practicals spanned PCR, gel electrophoresis, microscopy, plant anatomy, and ecological sampling.

Gold Medal.

Top 40 and Gold Medal.

Bronze and Silver Medal.