Asher Arya

A language built for Void Mining, a Unity game I'm making where players program their own autonomous mining bots.

• Engineered a full end-to-end execution pipeline for a custom statically typed language in C# by implementing a hand-written scanner, recursive-descent parser with operator precedence, and AST resolver/typer pass specified via a formal EBNF grammar, and interpreter.
• Resolved function call ambiguity and enforced type safety by engineering a two-phase overload resolution system that evaluates exact parameter matches before traversing a custom type-promotion lattice across nested lexical scopes.
• Built a spec-driven test harness in GitHub Actions that generates and runs 283 regression tests from JSON specifications.

GPU-scale Monte Carlo simulation used to guide state-space partitioning for Quasi-Markov Chain verification of autonomous controllers, fitting minimum volume enclosing ellipses to tighten upper bounds on failure probabilities.

• Engineered high-throughput CUDA C simulations that executed billions of Monte Carlo samples per second to evaluate rare failures.
• Tightened failure bounds by up to 30% by designing an enclosing ellipse algorithm combining bisection search with gradient descent.
• Accelerated neural network evaluation pipelines for zonotope reachability analysis by translating TensorFlow functions into native CUDA C and implementing custom kernels for eigendecomposition and backward-mode automatic differentiation.

A homelab I run for friends and family, hosting media, identity, and web services for 14 active users on a Proxmox hypervisor provisioned with Ansible and Terraform. This site is served from one of its containers.

• Orchestrated services for 14 active users by managing Proxmox LXC containers, firewall rules, and cluster config with Terraform, configuring services inside each container with Ansible, and enabling NVIDIA GPU passthrough for NVENC hardware transcoding.
• Centralized identity management across 5 downstream services by deploying Authentik as an identity provider for LDAP and OIDC.
• Enforced least-privilege network access across 12 containers and 2 nodes with default-deny policies and reusable Terraform security groups that allow-list only required ports, block lateral movement, and restrict admin access to Tailscale and known devices.

An action-roguelike set in a massive industrial complex consumed by four autonomous, corrupted AI factories. You play successive iterations of a robot built by the Mother of Preservation, sent down to destroy the Mothers of Dispersion, Recycling, and Propagation. Salvaged scrap is sent back up to permanently upgrade your factory between runs.

• Implemented the core engine systems for an action-roguelike, such as the ECS registry, scene management, procedural map generation, and a data-driven NPC dialogue system.
• Enabled accurate hit detection for non-rectangular projectiles and enemies by building a physics and collision pipeline with semi-implicit Euler integration, quadratic air drag, AABB broad-phase checks, narrow-phase tests, and minimum-overlap resolution.
• Reduced per-frame processing from 40ms to 10ms by adding cached filtered component views to the ECS registry, kept in sync through an observer pattern, allowing the collision, AI, physics, and rendering systems to iterate only over their relevant entities.

LLM planners paired with probabilistic model checking to give formal guarantees for navigation under uncertainty, outperforming a reinforcement learning baseline across diverse grid-world configurations.

• Architected a closed-loop verification pipeline that intercepts LLM-generated execution plans, routes them through a formal verifier for safety checking, and prompts the model to self-repair failing logic, resulting in a first-author FSE '26 companion publication.
• Demonstrated that verifier feedback scales with model size and outperforms prompt repetition by benchmarking five feedback mechanisms across three model sizes in a batched OpenAI/Gemini API harness with rate limiting, response caching, and seeded runs.
• Cut LLM output tokens by up to 73% while meeting 20% more requirements by re-architecting the system to synthesize rules over the state space instead of per-state plans and refine them with worst-case analysis, decoupling rule-set size from environment size.

• Ran weekly Scrum ceremonies, sprint reviews, and demo-day evaluations for 6 student teams shipping software for industry and clinical partners (SickKids, Korotu), assessing progress against user stories, technical depth, and deployment readiness.
• Drove SickKids' Echo Talk speech-therapy app to selection at the UofT Software Engineering Showcase by shaping per-user Azure speech models for dysarthria and stuttering, giving iterative feedback across four deliverables, and cutting scope to a shippable MVP.

• Streamlined the digital storefront experience by developing custom CSS templates and writing client-side JavaScript to dynamically render promotional cards within collection product grids during a major migration to Shopify.
• Cut manual product cataloguing by over 90% by deploying a TypeScript (Node.js) and PostgreSQL microservice on Render that used the Shopify Admin API to group and retitle newly imported products by populating custom metafields.
• Saved 10+ hours per week of manual merging by engineering a GitHub Actions CI/CD workflow that synced 6 designers' Shopify theme branches, replacing Shopify's manual copy-paste process with one-click merges and production releases.

• Cut weekly ticket processing time by 15% by developing and shipping an internal issue-tracking service in Microsoft Power Apps, iterating on direct feedback from executive leadership to eliminate operational bottlenecks, such as building a Teams integration.
• Built a React dashboard that polls the internal pricing database API to chart and monitor live energy prices for investors.
• Accelerated feature deployment cycles and reduced production runtime errors by engineering automated Jest/React testing suites and refactoring SQL query logic to optimize data retrieval pipelines across enterprise energy management tools.

• Courses: Automated Reasoning with Machine Learning (A+), Neural Networks and Deep Learning (A+), Software Engineering for Machine Learning (A+).

• Courses: DSA (96%), Numerical Analysis (91%), Artificial Intelligence (93%), Intelligent Systems (96%), Software Construction and Object-Oriented Programming (97%), Applied Machine Learning (94%).
• Awards and Recognitions: Science Scholar Dean's Honour List (min. 90% avg.), Dean of Science Scholarship (awarded by faculty), Charles & Jane Banks Scholarship (faculty recommendation), Trek Excellence Scholarship for Continuing Students (top 5% of year)