NICHOLAS FIDALGO

Bentonville, AR

Harvard University

M.S. Computer Science
B.A. Computer Science (Honors)
I am originally from Bentonville, Arkansas, now attending Harvard University in Cambridge, Massachusetts. I am an incoming senior graduating in May 2026. I live in Kirkland House and while on-campus I participated in Harvard Tech for Social Good and Harvard Data Analytics Group. I am earning a Bachelor of Arts in Computer Science with Honors and a Master of Science in Computer Science concurrently. Prior to university, I was an avid violist, MOSTEC 2021 scholar, and Hispanic Scholarship Fund and Youth Leadership Institute 2021 scholar.

About Me

NICHOLAS FIDALGO

I have experience in software engineering, data science, and entrepreneurship.

Netflix

• I will be joining the Playback Data Systems team this summer as a Software Engineer Intern in Los Gatos, CA.

Experience

Software Engineer Intern | Los Gatos, CA | June 2025 - August 2025

Massachusetts General Hospital

Data Scientist Intern | Boston, MA | August 2023 - July 2024

• Applied PCA, z-score standardization, and KNN imputation reducing data dimensionality by 86% and retaining 90% of variance, accelerating model training by 300%.

• Developed 5 custom features boosting polynomial regression model accuracy by 70% up to 94% for predicting MRI lesion counts, enabling insights into GLP-1 agonists' effects on multiple sclerosis.

• Automated clinical data processing pipeline for 49 patient records from MGH's registry, integrating Python and SQL, co-authoring a peer-reviewed publication.

Brill

Co-Founder | Cambridge, MA | January 2025 – Present

• Co-founded an AI-powered SAT preparation platform offering personalized tutoring, adaptive practice, full-length Digital SAT simulations, serving 10+ paying customers with a 4.9/5 user rating.

• Developed full-stack web application using React, Next.sjs, Supabase, and Stripe to build dynamic, responsive interfaces with LaTeX math rendering, user authentication, payment processing, and multi-table statistics tracking.

• Integrated OpenAI and Gemini APIs to generate and validate 2,000+ questions and curate 15+ practice tests.

Integrated two novel ML-based enhancements to LSM-trees: (1) a classifier to skip unnecessary Bloom filter checks achieving up to 2.28x GET latency reduction and (2) learned Bloom filters reducing memory footprint by 70–80%.

Learned LSM-Trees with Predictive Filtering

Here are my projects on algorithms and data systems research, AI, and ML projects.

Projects

NICHOLAS FIDALGO

Simulating Evolvability as a Learning Algorithm

Implemented first large-scale empirical investigation of evolvability to simulate evolution as a constrained learning process across 6 Boolean function classes and first exploration of majority and general conjunctions/disjunctions. Discovered majority and general conjunctions/disjunctions can evolve under aggregate-only feedback (previously unknown in theory), validated theory knowledge such as evolvability in monotone case and non-evolvability of parity. First empirical investigation of different distributions, neutral mutation disallowance, and initialization constraints.

Scalable LSM-Tree Storage Engine

Engineered high-performance LSM-tree in C++ using skip list memtable, Bloom filters with variable FPR, fence pointers, achieving sub-linear latency scaling from 100MB to 10GB datasets and up to 40% improved write throughput. Implemented novel hybrid compaction strategy combining tiering, lazy leveling, and leveling maintaining throughput under write-heavy workloads and improvements in 80/20 Zipfian skewed key and query distributions. Achieved near-linear scalability up to 16 threads and 32 concurrent clients, delivering 12× latency reduction and 25× throughput increase compared to single-threaded baselines, testing 8-dimensions of evaluation.

Forensic Sketch AI

Designed and implemented a multimodal pipeline integrating LoRA fine-tuned CLIP and Stable Diffusion v1.5, achieving 21% higher SSIM and 25% higher PSNR over baseline models in forensic sketch generation. Fine-tuned CLIP using LoRA on self- and cross-attention layers, improving text-sketch alignment by 9% and reducing perceptual error (LPIPS) over iterations validated through ablation studies. Developed iterative refinement process with dynamic embedding updates, enhancing quality across 5 refinement cycles.

Enhanced U-Net for Semantic Segmentation

Engineered enhanced U-Net architectures with residual blocks, batch normalization, and dilated convolutions, improving segmentation accuracy by 3.6–4.7% over the baseline model on urban scene datasets. Developed hybrid U-Net-FCN model with skip connections, achieving 0.93 Dice for sky and 0.91 for road segmentation. Optimized loss functions using a combined Dice and Cross-Entropy approach, mitigating class imbalance and improving segmentation accuracy for irregular objects like sidewalks (18%) and buildings (37%).

NICHOLAS FIDALGO

Skills