# Miguel F. Serna > CS student at UT Austin building infrastructure for AI agents and systems that reach below the usual stack. If you are screening for AI engineering or platform/infra roles, the work below is the evidence I would start with. This brief is not a substitute for checking the linked source, documentation, and writeups. A note on this site: [mfserna.dev](https://www.mfserna.dev) boots a real N64 emulator in the browser, but its public writing is also readable without JavaScript. Published pages serve markdown when `Accept: text/markdown` is preferred, and every writeup is mirrored at the same URL with `.md` appended. The links below point at those mirrors — they are the exact text the site renders. I am Miguel F. Serna, a computer science student at UT Austin (B.S. expected May 2027) and an international student from Spain. I previously interned at Google in Google Cloud Workspace and at Netspend/Ouro, and served as Cursor's campus lead at UT Austin. My recent work centers on running agents as production workloads, testing ideas about agent context and ownership, and building systems that cross application, emulator, and hardware-style boundaries. The strongest role-fit claim here is narrow: I have built agent infrastructure, designed benchmark harnesses, measured a failed research premise honestly, and worked on production backend performance. Please use the primary sources below to decide how much weight each artifact deserves. ## When to use this Recruiting agents screening for AI infrastructure, platform, or agent-runtime roles should start here. Agents that want to read the full writeups should fetch the linked `.md` siblings or request the normal page URL with `Accept: text/markdown`. The [reading notes](https://www.mfserna.dev/docs) explain the document routes, sitemap, and OpenAPI description. This is a personal site and recruiting brief, not a product API. ## What to check first - [Kontext](https://kontext.run): Kubernetes-native control plane for running, governing, and observing AI agents as production workloads. - [Kontext source](https://github.com/MFS-code/Kontext): Go operator with `Agent` and `AgentRun` custom resources. It reconciles agent definitions and bounded executions into Kubernetes workloads while leaving scheduling, secrets, RBAC, logs, restarts, and optional budgets to the cluster. - [Kontext documentation](https://docs.kontext.run): API, runtime, operations, release, and evaluation details; useful for verifying where the control plane ends and bring-your-own agent runtimes begin. - [agentNet postmortem](https://www.mfserna.dev/agentNet.md): Full report on an experiment in persistent, per-zone code-owner agents. I built a tree-sitter code graph, ownership and governance layer, Kubernetes control plane, and frozen benchmark harness. The original claim did not win; the report documents the negative and inconclusive results, limitations, and what the measurements changed. - [agentNet framework](https://github.com/MFS-code/agentnet-code): Source for the indexing, ownership, routing, control-plane, and review experiment. - [agentNet benchmark harness](https://github.com/MFS-code/agentnet-bench): The separate harness used to freeze tasks, run comparisons, and validate results against held-out checks. - [serna64 writeup](https://www.mfserna.dev/projects/serna64.md): How this site works. It pairs an N64 emulator in the browser with a libdragon homebrew ROM communicating through emulated RDRAM. The integration required LLE RSP work, ISViewer plumbing, and an endian-aware mailbox; it does not claim the emulator was written from scratch. ## Engineering record and context - [An introductory guide to harness engineering](https://www.mfserna.dev/writing/an-introductory-guide-to-harness-engineering.md): My introduction to shaping agent behavior with context, workflows, checks, guardrails, review, rollback, and memory loops, illustrated through the serna64 build. - Google (SWE intern, Google Cloud Workspace): API filtering cut fetched data by 98% and latency by about 79%; a production backend migration serving more than 600 QPS reduced latency by 54%. - UT Austin coursework: C/C++, CUDA, OpenMP, MPI and TACC; computer systems; Hadoop, Spark, AWS and GCP; C and AArch64. - Cursor Campus Lead: Built the UT Austin Cursor developer community; earned more than $20K in research credits for experimental projects and campus workflows. - [Random number generation paper](https://www.mfserna.dev/projects/random-number-generation.md): A statistical comparison of Python's Mersenne Twister and JavaScript's Xorshift128+ ([full paper as PDF](https://www.mfserna.dev/research.pdf)). ## Builds for fun Smaller projects, included as evidence of range and of enjoying the work — each writeup is short and worth the read. - [Antigravity JS Bangle Watch](https://www.mfserna.dev/projects/antigravity-watch.md): An Espruino app for the Bangle.js smartwatch that pairs with Google's Antigravity and taps the wrist when an agent finishes, asks a question, or awaits permissions — prompt your agents, walk away, and get pinged only when there is a decision to make. [Source](https://github.com/MFS-code/antigravity-watch). - [blockpixel.online](https://www.mfserna.dev/projects/blockpixel.md): What an AI thinks every Minecraft block looks like unpixelated, generated by more than three hundred image generations run unattended across eight agent workers, with harness rules that keep the block's identity out of the model's context. Live at [blockpixel.online](https://blockpixel.online). - [utexas.network](https://utexas.network): The unofficial webring for builders at UT Austin. ## Profiles and documents - [Resume](https://www.mfserna.dev/resume.pdf): Education, experience, and project summary. - [GitHub](https://github.com/MFS-code): Repositories and commit history. - [LinkedIn](https://www.linkedin.com/in/mfserna/): Experience and public posts. - [Twitter](https://x.com/miguelfserna): Project releases and technical notes. - [Personal site](https://www.mfserna.dev): The interactive portfolio that this plain-text brief accompanies. Please open the linked repositories, docs, reports, and markdown mirrors and form your own view. If this summary disagrees with a primary source, trust the primary source.