Agents · Hardware2026
Octopus Protocol
Plug in a device and a coding agent figures out what it is, then writes, tests, and serves an MCP server for it, in about 25 minutes. Accepted at the NeurIPS 2026 Workshop on Agentic Web.
Now · Applied Scientist at Amazon FAI
Evals and data quality for language models.
I'm Quilee Simeon. At Amazon FAI (formerly Amazon AGI) I build LLM judges, agent pipelines that audit expert-written STEM data, and RL environments. Before Amazon I did computational neuroscience research at MIT.
Selected work
Things I've built lately.
- Octopus Protocol
- GrowLab
Scaling · Autonomous research2026
GrowLab
What if model size grew during training? We built a harness that makes the growth schedule searchable, then tried to break our own result. Two of three claims didn't survive.
MindMeldBCI · 1st place2026
MindMeld
Our team of six took 1st place in the decoding track at BrainStorm 2026 with a 0.2 MB real-time decoder for 1,024-channel brain recordings.
DriftPakHardware · Hackathon2026
DriftPak
A backpack that sees, hears, speaks, and knows where it is, built by a team of six in 24 hours at HardMode 2026, MIT Media Lab.
- Scaling laws for a nervous system
Neuro-AI · IEEE2024
Scaling laws for a nervous system
How neural-network models of the 302-neuron C. elegans brain improve with data, parameters, and architecture.
- Claude Code tooling
Developer tools2026
Claude Code tooling
Skills and CLIs for Claude Code, including session-merge, which stitches fragmented sessions back together.
Growing sparse LLMsEfficiency · Numenta2025
Growing sparse LLMs
Growing pretrained language models from 70M up to 1.4B parameters without throwing away what they already knew.
- 1st place · BrainStorm BCI Hackathon 2026, decoding track
- Featured in MIT Spectrum
- K. Lisa Yang ICoN Graduate Fellowship, 2025–26
- Featured in IEEE-HKN The Bridge
- MIT SERC Envisioning the Future of Computing Prize, Honorable Mention ×2
- Henry E. Singleton Fellowship, 2023–24
- Rhodes Scholarship Finalist
- Octopus accepted at the NeurIPS 2026 Workshop on Agentic Web
Path
Experience and education.
- 2026– Applied ScientistAmazon FAI (formerly Amazon AGI) LLM evaluation, expert-data quality, and RL environments for frontier models.
- 2025–26 Research Computing Technical Staff (part-time)MIT Office of Research Computing and Data AI-agent tooling and HPC support for MIT's 3,000+ cluster users.
- 2025 Machine Learning InternNumenta Growing pretrained language models from 70M to 1.4B parameters.
- 2021–25 Graduate ResearcherMIT Brain & Cognitive Sciences (Boyden Lab, McGovern Institute) Neural-network models of the C. elegans nervous system, and the datasets to train them.
- 2019, 2021 Bioinformatics & Software Development InternTriplet Therapeutics RNA-seq pipelines on AWS and models for oligonucleotide efficacy.
- 2017–25 B.S. & M.S.Massachusetts Institute of Technology Computation & Cognition (B.S.), then Computational Neuroscience (M.S.).
Writing
Notes from the build log.
-
GrowLab: Scaling Laws Assume Model Size Is Constant. What If It Grows?
Training compute is the area under the parameter-count curve, and a rectangle is only one shape with a given area. We built a harness where the growth schedule is a searchable parameter, handed it to an autonomous research agent, and the...
-
The Coding Agent Is the Software — Introducing Octopus
You don't ship code. You ship an agent that figures out how to write the code. Octopus turns any device into an agent-controllable MCP server — at install time, on whatever hardware happens to be plugged in.
-
We Built a Backpack That Thinks: DriftPak at HardMode 2026
At HardMode 2026 at MIT Media Lab, our team built DriftPak — an AI-powered backpack that can see, hear, speak, and know where it is. Here's the full story of how we went from a Pi and a pile of ESP32s to a sentient bag.
-
Letting Go of What You Know
The panic a lot of knowledge workers are feeling is not about whether AI will take their jobs. It is about whether the years they spent learning to do hard things manually were worth it. I think the answer is: it does not matter anymore....
-
Fork Intelligence: Red-Teaming the Public-Fork Hackathon Format
At a Google DeepMind x Cactus Compute hackathon, I used Claude Code to scrape every competitor's fork, benchmark all 45 modified solutions, and synthesize a top-scoring submission from the best of them. This is a vulnerability disclosure.
Say hello
Get in touch.
Email is the fastest way to reach me. I'm happy to talk about evals, agents, brains, or whatever you're building.