Edjutawee Dawit

Experience

Purdue University

Nov 2025 – Present

Undergraduate Researcher (Prof. Abulhair Saparov)

Researching curriculum learning and reasoning-depth limits in transformers on synthetic graph-search tasks.
Found a reproducible one-hop reasoning plateau under uniform training, while curriculum learning reaches near-perfect deeper-search performance.

Microsoft

May 2026 – Aug 2026

Software Engineer Intern, Azure Applied AI

Shipped SHAP explanations and model monitoring (PSI/KS drift checks) for a production XGBoost risk platform, combining label-free performance estimates and fail-safe reports.
Designed and built a 15-case, five-metric evaluation harness for risk accuracy, SHAP faithfulness, groundedness, feature overlap, and offline F1, backed by 200+ automated tests.
Integrated grounded LLM narratives into HTTP and Azure Storage Queue inference paths; isolated contextualization failures with bounded-latency fallback.
Provisioned private, keyless Azure OpenAI with Bicep and Entra ID/managed identity across serving, infrastructure, and failure paths.

Eli Lilly

Nov 2025 – May 2026

LLM Researcher

Built an LLM-driven cybersecurity intelligence system; probed robustness under adversarial prompt perturbations and conflicting evidence injection, surfacing interpretability failure modes.
Engineered a modular Python framework leveraging LLMs to enrich cybersecurity indicators with automated tagging, contextual narratives, and inter-indicator relationship mapping.

NASA Beyond the Algorithm Challenge

May 2025 – Aug 2025

Software Engineer Intern / ML Engineer

Built a PyTorch→ONNX→TensorRT pipeline achieving 0.92 IoU and 10× CPU speedup for real-time SAR flood detection; evaluated degradation under distribution shift.
Developed a real-time SAR image processing pipeline for Sentinel-1 satellite tiles, achieving 250ms inference latency and enabling 5% faster flood detection response.

Purdue University

Jan 2025 – May 2025

Computational / Backend Engineer

Engineered a fully autonomous AI research assistant screening 2,000+ academic papers using Selenium, GPT-4 API, and local Ollama models—reducing faculty workload by 95%.
Built a distributed document processing pipeline using Python multiprocessing + Redis, achieving 100+ papers/hour throughput across 2,000+ academic papers with 95% automation accuracy.
Engineered a fault-tolerant system with checkpointing and idempotent job processing, deployed on AWS EC2 spot instances with Prometheus/Grafana monitoring tracking p95 latency.

CorpusKey

Aug 2024 – Oct 2024

Full-Stack Software Engineer Intern

Designed an AI-powered lesson planning platform serving 500+ educators with React + Flask microservices, achieving 99.5% uptime and a 70% reduction in planning time.
Implemented ChatGPT API integration with rate limiting and caching; optimized PostgreSQL queries with indexing strategies for sub-100ms response times.

Independent AI Researcher

Winter 2024

Reinforcement Learning Chess Agent

Developed a reinforcement learning chess agent trained via self-play and policy gradients, evaluated against Stockfish.
Automated gameplay via Selenium, training the agent to 1800 ELO (top 15% of players) in 500+ self-play episodes.

John P Stevens

Spring 2024

Machine Learning Researcher

Conducted hyperparameter sweeps and ablation studies comparing SVM, Gradient Boosting, LSTM, and ANN models across 3 real-world datasets.
Cross-validated with 50+ academic sources and implemented reproducible baselines from Hugging Face and PyTorch.

Temple University

Jun 2023 – Aug 2023

Computer Vision Researcher

Researched CNN optimization for image recognition via transfer learning and data augmentation; improved model accuracy and generalization on classification benchmarks.

Publications & Research Projects

Algorithmic Generalization in Deep Transformers

In Submission to NeurIPS 2026

Used scratchpad tokens as a mechanistic probe; Fourier analysis revealed sequential computation steps—not parameter capacity—are the true bottleneck for grokking.

Backdoor Trigger Detection | Jane Street Dormant LLM Puzzle

2026

Probed hidden triggers in three 671B-parameter DeepSeek-V3 fine-tunes via black-box probing and activation-layer SNR analysis across 13 attention layers.
Submitted methodology and findings to the $50k AI safety prize.

Two-Phase Scaling Law for Neural Chess Search

Submitted to IEEE CoG 2026

Designed and parallelized a fixed-anchor tournament of 7,000+ games across an HPC cluster to evaluate Stockfish 16 (NNUE) depth scaling.
Validated a two-phase law (R²=0.984, Chow test F=19.02), quantifying ~66% per-ply gain reduction vs. classical engines.

Mechanistic Interpretability of a DQN Chess Agent

2024 – 2025

Identified neuron clusters encoding attack/defense/king-safety inside an RL chess agent trained via self-play and policy gradients to 1800 ELO.
Discovered phase transitions in internal representations corresponding to brittle value estimation and policy overfitting.

RL for SPY Trading

Accepted as Full Paper to IEEE IDS 2025 (FinRLFM Track)

Demonstrated DQN/Double DQN outperform SAC/A2C in noisy market environments.
Integrated LLM-based sentiment signals via DeepSeek, improving Sharpe Ratio metrics.

Edjutawee Dawit

Education

Experience

Publications & Research Projects

Engineering Projects & Open Source

Technical Skills