Eric Solak | Founder of Alloy · Software Engineer

About

I build and lead software systems end-to-end, from architecture and implementation through deployment.

Through Alloy, I’ve worked directly with business owners across Canada to scope projects, translate their needs into software, and deliver production systems myself.

At CATTLElytics, I also led a team of engineers through the development of a production AI product. I’ve learned how to communicate effectively with non-technical stakeholders while still being able to go deep technically and lead engineering work.

Get in touch Visit Alloy

Experience

Jun 2026 — PresentToronto, ON

Alloy

Founder

Python · Claude Code · React · TypeScript · JavaScript · HTML/CSS · Cloudflare Workers · REST APIs

Grew Alloy into a solo consultancy whose automations save client businesses 100+ hours of manual work per month, handling all client-facing and technical work single-handedly.
Built fully and semi-autonomous AI workflows automating invoicing, document summarization, sales tracking, and commission calculations for small and mid-sized businesses.
Delivered automation and website projects for businesses across Canada, owning requirements gathering, scoping, build, and handoff within a 1–3 week window per project.

Design and build production systems end to end: custom sites in React and TypeScript on Cloudflare Workers, plus fully and semi-autonomous AI automation pipelines fitted to each business’s existing systems.
Architected integration layers that plug automations into the tools clients already run on, from Microsoft 365 and Google Workspace to QuickBooks, Xero, Sage, Salesforce, Shopify, Stripe, Slack, Airtable, and Notion, handling authentication, data mapping, and error recovery across disparate systems.
Validated automations against live client data before handoff, iterating on edge cases found through real usage.
Sustained an 80%+ client retention and referral rate through direct, hands-on relationship management.

Automate your current systems

…

Visit alloyconsulting.net

Sep 2025 — Apr 2026Hamilton, ON

CATTLElytics Inc.

Industry Capstone Project

Python · Flask · OpenAI Whisper · React Native · Expo · SQLite · REST APIs

Led AI development for Pocket AI, a voice-driven task management app supporting dairy operations for 1 in 50 cows across North America.
Directed a team of 5 engineers through the full product lifecycle, from requirements and architecture to production handoff, delivering an app being integrated into CATTLElytics’ platform.
Integrated OpenAI Whisper for real-time audio transcription, achieving 98% voice recognition accuracy under simulated barn noise, exceeding the 95% project target.

Built a Flask backend with REST APIs for authentication, task lifecycle management, and sync with CATTLElytics’ proprietary storage system.
Implemented an offline-first architecture with local SQLite storage and background syncing for reliable performance in low-connectivity farm environments.
Designed and built mobile UI components in React Native (Expo), including task views, audio capture flows, and speech confirmation feedback; usability testing yielded a mean satisfaction score of 4.4/5.

2022 — May 2026Hamilton, ON

McMaster University

Bachelor of Engineering, Software Engineering

cGPA 3.70 / 4.0

AI / Machine Learning: Applications of Machine Learning, Data Mining, Probability & Statistics
Software Engineering: Software Design IV, Software Testing, Human-Computer Interfaces
Systems: Signals and Systems, Computer Networks & Security, Linear Optimization

Selected projects

01 / LLM built from scratch

musicGPT

Jul 2026 · Python · PyTorch · NumPy · tiktoken · pytest

A 45M-parameter GPT built from scratch in PyTorch (attention, RoPE, training loop, all hand-written), pretrained on ~1 billion Wikipedia tokens on a single GPU to write fluent, music-literate prose.

Live demo Code Training results

Implemented a decoder-only transformer (44.6M parameters — 6 layers, 8 heads, 512-dim) with every core component written by hand: multi-head causal self-attention, rotary positional embeddings (RoPE), LayerNorm with fp32 statistics, and GPT-style weight init — no pre-built transformer layers.
Engineered a streaming data pipeline that pulls official Wikipedia dumps over HTTP, strips nested wiki markup with a hand-rolled parser, and packs 0.92B GPT-2 BPE tokens into memory-mapped binaries — upweighting music articles 3× for domain fluency.
Trained at the compute-optimal ~20 tokens-per-parameter budget on a single L4 GPU with mixed precision, cosine LR scheduling, and gradient accumulation — sustaining ~90k tokens/sec and reaching validation perplexity ≈ 26 with train and val loss tracking tightly (3.24 / 3.26): no overfitting.
Built an 11-test pytest suite: manual-attention vs. fused-kernel equivalence, causality, RoPE rotation geometry, and an init-loss sanity check that caught a real weight-tying bug during development.

From raw dump to working language model

Wikipedia dump · streamed over HTTP · 0.92B tokens

Causal Self-Attentionwith RoPE, hand-written

LayerNormmanual, fp32 stats

Feed-ForwardGELU, 4× expansion

Inside each of 6 transformer blocks — every layer from scratch

Training loop · mixed precision · cosine LR · ~90k tokens/sec

Final model · val perplexity ≈ 26 · coherent encyclopedic prose

02 / CATTLElytics Inc.

Pocket AI

Sep 2025 – Apr 2026 · Python · Flask · React Native · OpenAI Whisper · REST API

Project highlighted on official McMaster platforms

Speak naturally, get structured tasks. End-to-end voice-to-task system built with Whisper, an LLM pipeline, Flask API, and React Native.

Code

Integrated OpenAI Whisper for real-time audio transcription, validated through usability testing that achieved 98% system voice recognition accuracy under simulated barn noise, exceeding the 95% project target.
Built a Flask backend powering voice-to-task workflows with REST APIs for authentication, task lifecycle management, and synchronization with CATTLElytics' proprietary storage system.
Implemented an offline-first architecture with local SQLite storage and background syncing, ensuring reliable performance in low-connectivity farm environments.
Designed and built mobile UI components in React Native (Expo), including task views, audio capture flows, and speech confirmation feedback; usability testing yielded a mean satisfaction score of 4.4/5.

03 / Machine learning

Wildfire Risk Classifier

Sep 2025 – Jun 2026 · Python · PyTorch · FastAPI · Docker

Neural network trained on 118,000+ NASA satellite samples. Classifies wildfire risk with 94.3% accuracy, beating every baseline we tested — now deployed as a live REST API.

Live API Interactive map Report Code

Deployed a production REST API using FastAPI and Docker on Render, serving real-time wildfire risk predictions from a containerized service.
Developed a neural network to classify regions into low, medium, and high wildfire risk using 118,000+ samples from NASA FIRMS and Open-Meteo datasets.
Designed a 4-layer MLP with batch normalization, dropout regularization, and weighted cross-entropy loss to address class imbalance across risk categories.
Achieved 94.3% test accuracy with high F1-scores across all classes, outperforming Random Forest and SVM baselines by 15%.

··Custom model

04 / Computer vision

SneakPeek

Jan 2025 – Mar 2025 · Python · PyTorch · React Native · JavaScript

Point your phone at any sneaker and get an instant AI identification plus purchase links. Custom DenseNet201 model, multi-agent Blackboard architecture, React Native app.

Code

Built a full-stack mobile app for AI-powered sneaker identification and social forum, where users upload images and receive an AI-identified shoe model, brand, and purchase links.
Trained a custom DenseNet201 model on a labelled dataset, applying a data augmentation pipeline (perspective transforms, saturation jitter, geometric transforms) to maximize generalization.
Architected a multi-agent Blackboard system where Google Gemini 2.0 Flash, GPT-4o, and the custom PyTorch model each submit predictions to a shared space; a central Gemini controller arbitrates and selects the final result.

How a prediction is made

Sneaker photo uploaded

Gemini
2.0 Flash

GPT-4o

DenseNet201custom PyTorch

3 agents predict in parallel

Blackboard · shared prediction space

Gemini controller arbitrates → best match

Brand, model & purchase links

Drone searching a procedurally generated island

05 / Autonomous systems

Autonomous Drone Rescue

Sep 2023 – Dec 2023 · Java · JUnit · Maven · Git

Autonomous flight software that navigates procedurally generated islands to locate and rescue targets. Built in Java with a team of three.

Code

Developed autonomous flight software to locate survivors across unknown terrain, implementing algorithms for optimal pathfinding to systematically cover the area.
Designed and executed test cases with JUnit to verify functionality of software components across a three-person team.

Skills

What I have worked with in practice.

Languages

Python

Java

C++

JavaScript

HTML/CSS

TypeScript

C

SQL

Frameworks & Libraries

PyTorch

FastAPI

Flask

React Native

NumPy

Pandas

scikit-learn

Tools

Docker

Render

Claude Code

GitHub Actions

JUnit

SQLite

Contact

Let’s build something great together.

Send an Email solakeric28@gmail.com

Visit Alloy LinkedIn GitHub

Toronto, ON, Canada