00 - ryan polasky
north dakota · cdt
software engineer · backend + ai / llm tooling
ut dallas cs grad (may '26). two summers at apple. most recently built autopsy - forensic memory for ai coding agents.
01 - about
i'm a computer science graduate from ut dallas ('26). i spent my last two summers as a software engineering intern at apple, building out ai tooling for the wireless tech (WTE) & the camera & photos organizations.
i enjoy backend, devops, and making llms actually do what you want them to do. i also have a bad habit of rebuilding this website. this is version eleven, but the older ones are still around if you want to see them.
languages
data
ai / llms
devops
02 - experience
01.
customer engineering · remote · may 2023 → present
worked remotely alongside school, sustaining a 6.87% escalation rate well below department average through independent research and advanced troubleshooting. 0.27 min average call wrap time, 15% above department-average csat, and 2× applecare excellence award recipient (2024, 2025).
+15%
csat vs dept avg
02.
student leadership · dallas · may 2025 → may 2026
served as development lead within acm at ut dallas. mentored 1,200+ hackutd participants and debugged critical issues for 60+ teams. led meteormate, a react, fastapi, and postgres roommate-matching platform. also mentored four underclassmen through internship preparation.
1,200+
hackers mentored
03.
intern · cupertino · may–aug 2025
built automated parsing for thousands of on-device camera and photos tests, integrated llms into the failure-triage pipeline, and shipped a unified bug-report generator that attached relevant failure artifacts.
4,800%
manual triage reduction
04.
intern · cupertino · jun–nov 2024
built an llm-powered chat interface and python package for translating natural language into sql and nosql queries. migrated internal services from java 11 to jdk 21, addressed security vulnerabilities, designed aws/gcp ci/cd pipelines, and optimized an ios log analyzer for real-time rendering.
nl → sql
schema-aware planner
05.
intern · grand forks · jun 2021 → jul 2022
repaired, imaged, and deployed hundreds of machines for a critical healthcare-system expansion. handled urgent hardware deliveries during peak demand and set the department intake record at 300+ devices.
300+
intake record
selected system
a black box recorder for ai coding agents.
a one-line install drops a typescript plugin into opencode that streams every tool call, edit, rejection, and prompt to a fastapi service. rejected runs flow through a deterministic classifier and an optional gemma enhancer, then land as nodes in a pgvector-backed failure knowledge graph: 9 node types, 8 edge types, all in postgres.
when a new task starts, autopsy embeds the description, runs ann search against past failures, and walks the graph (run to symptom to failure mode to fix pattern) with temporal decay and counter-evidence dampening. the most relevant warnings get injected straight into the agent's system prompt before it generates a single line of code.
every agent run starts from a clean slate. they forget last week's failure, drift the types again, skip the migration again, and burn api credits relearning lessons you already rejected. there's no recorder, no investigation, no safety briefing. we wanted to build the ntsb for coding agents.
the closed-loop demo runs end to end on a real agent runtime. agent fails, autopsy records, graph grows, next similar task gets a warning, agent fixes it on the first try. fully local stack with 165 service tests passing across event ingestion, finalizer, classifier, extractor, graph writer, and embeddings.
selected system
ai-driven roommate matching for utd students.
product lead
MeteorMate is built around the stuff that decides whether roommates work: sleep schedules, cleanliness, guests, study habits, social energy, and lifestyle preferences. I led the project shape from roadmap to task breakdown so the team could move from broad housing idea to a concrete matching product.
I defined and documented the end-to-end shape of the system: a Next.js/React interface for students, a Python/FastAPI backend for matching workflows, Postgres for structured application data, and Firebase where it made sense for auth and realtime product needs.
A big part of the work was making the service reliably boot and route in the environments the team actually used. I drove fixes around serverless-style entrypoints, routing behavior, and local parity so the backend could run predictably instead of only working on one person’s machine.
selected system
a 2D survivor roguelike about a druid fighting capital encroachment.
four forms, each shaped to counter a specific boss archetype. fox bleed whittles the brute, bear smashes through armor, frog zones the bullet-hell phase, druid handles everything in between. swapping mid-fight builds a combo multiplier, so the meta-loop is "stay alive long enough to learn the matchup" rather than "pick a main and grind."
each boss is a different system stress-test. the lumberjack drops trees that permanently rewrite the navmesh. the foreman pours liquid concrete that hardens into area-denial. the contractor wraps himself in vehicle armor. the CEO shrinks the arena by paving it over while you fight him - lose ground, lose footing.
selected system
a cache-first 3d record room built from my discogs collection.
the centerpiece is a procedural 3D shelf filled with the actual albums i own. each record is interactive: pull it out of the slot, flip to the back jacket, see the tracklist and the description load into an info panel. a turntable holds whatever record is currently "active" at the center of the shelf. roomdressing fills the rest - plant, mug, book stack, wall art, a paper note from me on the floor.
selected system
end-to-end triage pipeline for 1000+ camera tests: parse, group, classify, file.
every cycle the camera org ran 1000+ on-device tests, and manual triage of the failures was the slowest part of the loop. the pipeline ingests raw xml results, parses failure signatures, groups related failures so similar problems collapse into a single investigation, and feeds each cluster into an LLM for root cause analysis before anything reaches a human.
integrated Gemini for bug deduplication and root-cause hypothesis surfacing. 75% accuracy on duplicate identification meant new failures could be matched against the existing radar history before a fresh ticket got filed, eliminating redundant radars and speeding the end-to-end triage workflow by 4,800%.
the back half of the pipeline writes bug reports automatically. multi-threaded creation with exponential-backoff retry handles transient radar failures, database-backed regression analysis ties new failures to known historical patterns, and logs + failure context attach themselves to each report. ran daily for 50+ engineers across the camera org.
selected system
natural-language query interface for internal data stores.
selected system
browser-only generator for animated github readme banners.
