Mathematics & Computer Science · Boston University

Ahmed Alwahedi

I build machine-learning systems and measure whether they work: reinforcement-learning agents, fine-tuned image models, multilingual retrieval and multi-agent LLM platforms.

Résumé Email GitHub LinkedIn
Portrait of Ahmed Alwahedi in a navy suit and light-blue tie.
Studying
BA Mathematics & Computer ScienceBoston University · May 2027
Internships
7X · MBZUAI · ADIA Lab · Astra Tech2025–2026, Dubai and Abu Dhabi
Record
3.74 GPADean’s List ×4 · ISNR 2026 CTF, 1st place

Featured projects

Each one has a full case study: what I built, how I measured it, and where it falls short.

Bar chart titled 'The same battles, about 900× faster than Pokémon Showdown'. Showdown 7.69k turns per second (1×); Kanto with the same text output 185–189k (24–25×); Kanto engine without text 6.80–7.04M (884–916×).

Independent project · 2026

Kanto & Indigo: fast Pokémon simulator and self-play agents

A Rust simulator that reproduces Pokémon Showdown's Gen 1 battles exactly, ~900× faster. Agents trained on it by self-play beat TaurosV0, Metamon's best standalone Gen 1 OU policy.

faster than Pokémon Showdown at replaying the same battles
~900×
score against TaurosV0, which its authors describe as Metamon's best standalone Gen 1 OU policy
60.2%
differences from Showdown over 81,496 test battles
0

Rust · Python / PyO3 · PyTorch · PPO self-play · Differential testing · Monte Carlo search

Game AI / search

Othello search engine and pruning ablation

An Othello engine in Java with four search shortcuts stacked together. It searches 15.5× fewer positions than plain alpha-beta search, yet loses to it in timed games; I measured why.

fewer positions searched than plain alpha-beta
15.5×
moves considered per position, down from 11.8
4.45
score against plain alpha-beta at equal time
42%

Java 8 · Alpha-beta / PVS · Transposition table · Multi-ProbCut · Optuna

LoRA image of a fern fox in pencil on paper.

Independent project · 2026

Style LoRA data ablation on FLUX.2 Klein

Six LoRAs teach FLUX.2 Klein an illustration style I designed, each from a different mix of reference and synthetic images. 252 images scored blind by a VLM judge.

style score out of 5, base model vs best LoRA
1.49 → 2.92
LoRAs, one per data mix, same training recipe
6
of 252 style scores within one point across two judge passes
89%

FLUX.2 Klein 4B · LoRA · AI Toolkit · vLLM · Qwen3.8-27B (judge) · NVIDIA GB10

Other projects

Research, tools and experiments.

All 7 projects
Left: planning success for MAE alone 0.65, MAE to DINO 0.63 and MAE to SigLIP 0.64, with a note that Li et al. reported 0.02 and 0.00. Right: stacked bars over 21 stitches; four coarse capabilities are preserved in all of them, distance to the door is degraded in 8, and alignment with the door is degraded in 13, inconclusive in 4 and preserved in 4.

Independent project

Beyond Latent Alignment: auditing world-model stitching

A paper draft that re-implements DINO-WM model stitching. A reported collapse did not replicate, and a per-capability audit found door skills degrading in stitches that plan normally.

2026 PyTorch · DINO-WM · DINOv2 · SigLIP · MAE

A living room with a grey sofa, glass coffee table, green striped rug, walnut sideboard and shelving, and a white armchair.

Independent project

Haven: room scan to fit-checked furnishing plan

An app that turns an iPhone room scan into a millimetre-accurate floor plan and a furnishing plan, checking every piece against the measured room. A prototype, tested on a sample scan.

2026 TypeScript · Cloudflare Workers + D1 · Apple RoomPlan

Independent project

Digital Terrarium: multi-agent RL survival world

One small neural network, shared by 30 agents, keeps 86.3% of them alive in a simulated survival world, against 62.0% for a hand-written strategy and 0% for random play.

2026 Python · NumPy · PyTorch

Experience

Internships and research.

  1. Jul – Aug 2026

    Dubai, UAE

    Forward Deployed Engineering Intern — 7X

    Built a config-driven platform of AI employees for four business departments, with deterministic governance around every LLM call.

    • Four department agents on Python, FastAPI and CrewAI, each with scoped MCP tools, knowledge bases, scheduled duties and role-gated chat.
    • RBAC, per-agent scopes, tool/token/time budgets, human approval gates, audit and replay.
    • MCP/Oracle tools and PostgreSQL telemetry, deployed via GitHub Actions to Azure App Service.

  2. Jun 2026

    Abu Dhabi, UAE

    Machine Learning Research Intern — MBZUAI · UGRIP

    When a team of LLM agents produces a wrong result, which agent caused it? Built testbeds where the answer is known by construction and scored black-box and white-box attribution methods against it.

    • Built two ground-truthed multi-agent LLM testbeds (a linear and a star pipeline) with single-fault injection, so the responsible agent is known by construction.
    • Compared black-box attribution (leave-one-out, counterfactual repair, LLM judges) with white-box signals (gradient×input saliency, contrastive logit ablation, TracIn-style influence); judges found where failures surfaced but not their first cause, at both 1.7B and 14B.
    • A blind, trace-only structural rule recovered the first cause in 6 of 6 fault episodes where the 14B judge got 1 of 6; ground truth also exposed a label leak that had inflated an earlier result to a perfect score.

  3. Jun – Aug 2025

    Abu Dhabi, UAE

    Machine Learning Intern — ADIA Lab

    Diabetes risk classification on the Pima Indians Diabetes dataset, from model selection to a tool clinicians can use.

    • Benchmarked 4+ classifiers and selected an F1-optimised Random Forest via Optuna tuning: F1 0.71, recall 0.80, accuracy 0.77.
    • Subgroup fairness analysis by age to reduce bias; SHAP identified age, pregnancies and insulin as the top predictive features.
    • Deployed an interactive Streamlit app that lets clinicians classify new patient records in real time.

  4. Jul – Aug 2025

    Abu Dhabi, UAE

    AI Intern — Astra Tech

    An end-to-end automated pipeline that produces a daily bilingual (English and Arabic) AI news podcast for business and leadership audiences.

    • Tavily API for news aggregation, multi-agent LLMs for script generation, and ElevenLabs text-to-speech for audio.
    • Bilingual pipeline validation to keep English and Arabic episodes consistent.

Hackathons & competitions

  1. 2026

    ISNR Capture the Flag — 1st place

  2. 2025

    Boston, MA

    MIT iQuHACK

    Built a quantum–classical solver for MaxCut (splitting a graph's nodes into two groups to cut as many edges as possible) with QAOA in Qiskit, and benchmarked it against classical optimisation.

Education & skills

Boston University

BA in Mathematics and Computer Science

Expected May 2027 · Boston, MA · GPA 3.74 / 4.00 · Dean’s List ×4

Relevant courses

Data Structures, Analysis of Algorithms, Artificial Intelligence, Data Science, Linear Algebra, Probability, Differential Calculus

Languages
Python, Java, SQL, C, JavaScript, Go, R
AI / ML
PyTorch, scikit-learn, NumPy, Hugging Face, LoRA fine-tuning, RAG
Backend / infra
FastAPI, MCP, PostgreSQL, Azure, Git