Open to work
Mira Castellan
I build the systems that serve large language models: fast inference servers, careful evaluation and the tools around them. For six years I have worked between research and production. I like small, clear code and measurements that people can trust.
- Role
- Senior Machine Learning Engineer at Tidewater AI
- Location
- Brooklyn, New York
- Experience
- 8 years, 3 employers
- Newest skill
- AI agents, from February 2026
- This page
- Updated by itself on November 8, 2026, from 7 public sources
For recruiters
Can Mira do the job that you hire for?
Give a job. You see what the job needs, what Mira can prove with public work, and what you must ask.
13 skills with public proof. Each one has a receipt.
A receipt shows where the proof is, when JobsWithPay found the skill in public work, and what companies pay for it now.
Main work now
Python
- First proof
- May 2020
- Latest proof
- October 2026
- torchwave/torchwave, 14 merged pull requestsGitHub
- servekit/servekit, 9 merged pull requestsGitHub
- evalcards, 41.3K downloads a monthPyPI
- and 4 more public items
Open jobs on JobsWithPay that name this skill
21,480 jobs
Posted pay, middle of the range: $165K to $240K
Main work now
PyTorch
- First proof
- August 2020
- Latest proof
- October 2026
- torchwave/torchwave, 14 merged pull requestsGitHub
- mira-sample/tinyllama-speculator-1b, 12.4K downloads a monthHugging Face
- mira-sample/grasp-det-small, 1,900 downloads a monthHugging Face
- and 2 more public items
Open jobs on JobsWithPay that name this skill
4,431 jobs
Posted pay, middle of the range: $180K to $259K
Main work now
Model evaluation
- First proof
- May 2020
- Latest proof
- October 2026
- evalcards, 41.3K downloads a monthPyPI
- evalcards, 512 starsGitHub
- When Do Evaluation Sets Go Stale? A Study of 40 Model Releases, 88 citationsSemantic Scholar
- and 1 more public item
Open jobs on JobsWithPay that name this skill
2,976 jobs
Posted pay, middle of the range: $180K to $260K
Main work now
CUDA
- First proof
- October 2021
- Latest proof
- November 2026
- servekit/servekit, 9 merged pull requestsGitHub
- tinyserve, 1,840 starsGitHub
- triton-kernels, 24 starsGitHub
Open jobs on JobsWithPay that name this skill
1,365 jobs
Posted pay, middle of the range: $190K to $260K
New in the last 12 months
AI agents
- First proof
- February 2026
- Latest proof
- November 2026
Open jobs on JobsWithPay that name this skill
2,434 jobs
Posted pay, middle of the range: $170K to $235K
Main work now
Rust
- First proof
- March 2023
- Latest proof
- November 2026
Open jobs on JobsWithPay that name this skill
4,031 jobs
Posted pay, middle of the range: $180K to $250K
Show 7 more skills
Main work now
vLLM
- First proof
- May 2022
- Latest proof
- October 2026
Open jobs on JobsWithPay that name this skill
412 jobs
Posted pay, middle of the range: $185K to $265K
Used before
Computer vision
- First proof
- 2022
- Latest proof
- 2022
- mira-sample/grasp-det-small, 1,900 downloads a monthHugging Face
- Grasp Detection with Few Labels in Warehouse Scenes, 61 citationsSemantic Scholar
Open jobs on JobsWithPay that name this skill
1,922 jobs
Posted pay, middle of the range: $165K to $240K
Main work now
SQL
- First proof
- September 2025
- Latest proof
- March 2026
Open jobs on JobsWithPay that name this skill
16,200 jobs
Posted pay, middle of the range: $140K to $205K
Main work now
Triton
- First proof
- October 2021
- Latest proof
- December 2025
- triton-kernels, 24 starsGitHub
Open jobs on JobsWithPay that name this skill
86 jobs
Main work now
LLMs
Found by JobsWithPay. Mira did not list it.
- Latest proof
- November 2026
- servekit/servekit, 9 merged pull requestsGitHub
- evalcards, 41.3K downloads a monthPyPI
- mira-sample/tinyllama-speculator-1b, 12.4K downloads a monthHugging Face
- and 5 more public items
Open jobs on JobsWithPay that name this skill
5,648 jobs
Posted pay, middle of the range: $180K to $263K
Main work now
Model inference
Found by JobsWithPay. Mira did not list it.
- First proof
- May 2022
- Latest proof
- November 2026
- servekit/servekit, 9 merged pull requestsGitHub
- tinyserve, 1,840 starsGitHub
- Speculative Batching: Lower Latency for Shared Inference Servers, 214 citationsSemantic Scholar
- and 1 more public item
Main work now
Machine learning
Found by JobsWithPay. Mira did not list it.
- First proof
- August 2020
- Latest proof
- October 2026
JobsWithPay read the public sources on November 8, 2026. The dates are from the public work that we found. New: the first proof is in the 12 months before that day. Main work now: the latest proof is in those 12 months. Used before: the latest proof is older. Job counts and pay are from the open job posts on JobsWithPay. Pay is the posted yearly pay in US dollars, and shows only when 20 or more jobs give it.
Stated by Mira. JobsWithPay found no public source to check these skills.
Kaggle: mirasample Not checked. JobsWithPay cannot check Kaggle.
How Mira’s skills changed
Each bar shows the years in which Mira used the skill. A solid bar is from public work that JobsWithPay read: code, pull requests, packages, models and papers. An outlined bar is from the job dates that Mira wrote.
New in the last 12 monthsThe first public proof is in the last 12 months
- AI agents1 public repository (310 stars), 9 merged pull requests in agentkit/agentkit
Main work nowPublic proof in the last 12 months
- Python3 public repositories (632 stars), 23 merged pull requests in torchwave/torchwave and 1 more project
- PyTorch2 public repositories (120 stars), 14 merged pull requests in torchwave/torchwave
- Model evaluation2 public repositories (608 stars), 1 package (41.3K downloads a month)
- CUDA2 public repositories (1,864 stars), 9 merged pull requests in servekit/servekit
- Rust1 public repository (1,840 stars), 5 merged pull requests in ferrox/ferrox-runtime
- vLLM9 merged pull requests in servekit/servekit
- SQL3 merged pull requests in duckling-db/duckling
- Triton1 public repository (24 stars)
- LLMs3 public repositories (2,539 stars), 9 merged pull requests in servekit/servekit
- Model inference1 public repository (1,840 stars), 9 merged pull requests in servekit/servekit
Used beforeNo public proof in the last 12 months
- Computer vision1 model on Hugging Face (1,900 downloads a month), 1 paper (61 citations)
- A public item with a date in that year
- Between the first and the latest public proof
- Stated by Mira, from job dates
How fast Mira learns a new skill
The time from the first public work with a skill to a pull request that a different team merged. JobsWithPay read both dates on GitHub and in the package sources.
AI agents
3 months
First repository in February 2026.
First pull request merged in agentkit/agentkit in May 2026.
Rust
3 months
First repository in March 2023.
First pull request merged in ferrox/ferrox-runtime in July 2023.
CUDA
6 months
First repository in October 2021.
First pull request merged in servekit/servekit in May 2022.
Three pieces of work, and how Mira did them.
A list of skills does not show judgment. Each file shows one problem, the part that Mira did, and the part that an AI tool did. Mira wrote these files.
2024, Tidewater AI
Faster answers from a shared model server
38%
lower median response time
The problem
Many users shared one GPU server, and short requests waited behind long ones.
What Mira did
I designed speculative batching, wrote the scheduler, and measured it on three GPU types.
What an AI tool did
An agent wrote the first version of the benchmark scripts. I wrote the scheduler myself.
Proof
2025, open source
Evaluation reports that a team can repeat
41,300
downloads a month
The problem
Each model release had a different evaluation report, so nobody could compare two releases.
What Mira did
I defined one report format, built the tool evalcards, and studied 40 model releases with it.
What an AI tool did
An agent wrote the web pages of the reports. I reviewed them and wrote the scoring code.
Proof
2026, open source
Tools that let an agent test a model server
9
pull requests merged by a different team
The problem
Agents could start a model server, but they could not check if it answered correctly.
What Mira did
I wrote an MCP server and the tests that show when an agent reads a result incorrectly.
What an AI tool did
An agent wrote most of the server code. I wrote the tests first and rejected 3 of its 12 changes.
Proof
How Mira works with AI tools
I write the design and the tests myself. An AI agent writes most of the first version of the code. I read each line before it is merged, and I measure each change before I trust it.
Written by Mira.
Checked by JobsWithPay in the public code of Mira
- 38%
of the commits in 2026 name an AI agent as co-author (233 of 612 commits). In 2025 it was 4%.
We read the co-author lines of public commits. We do not read private code. The count is for the repositories of Mira on GitHub.
Numbers that JobsWithPay checked
Mira did not type these numbers. We read them from the public source each day.
- 43merged pull requests in projects of other teamsSource: GitHub
- 47,300downloads a month of the packagesSource: PyPI and npm
- 402citations of 7 papersSource: ORCID and Semantic Scholar
- 2,690stars on the repositoriesSource: GitHub
The last 30 days
Public work. JobsWithPay updates this list each day. Last update: November 8, 2026.
- New paper: Draft Models That Know When to StoparXiv, 2026ORCID and Semantic Scholar
- Pushed to tinyserveGitHub
- New version of tinyservev0.9.0GitHub
- Pull request merged in torchwave/torchwaveFuse the rotary embedding into the attention kernelGitHub
- Pull request merged in servekit/servekitAdd chunked prefill to the batch schedulerGitHub
- New model on Hugging Face: mira-sample/tinyllama-speculator-1bHugging Face
- New version on PyPI: evalcards0.6.0PyPI
One year of public work
149 days with public work on GitHub, from November 9, 2025 to November 8, 2026.
From the contribution calendar that GitHub shows on the public profile of Mira. A darker day has more contributions.
Experience
Mar 2023 to Present
Senior Machine Learning Engineer
Tidewater AI, New York
- Lead the inference team of four engineers for the company's chat product.
- Cut the median response latency by 38% with speculative decoding and a new batch scheduler.
- Built the evaluation harness that the research team uses for every model release.
Jun 2020 to Feb 2023
Machine Learning Engineer
Kestrel Robotics, Boston
- Trained the vision models for warehouse robots, from data pipeline to deployment on the robot.
- Moved model training to a shared GPU cluster and cut the cost for each experiment by half.
Aug 2018 to May 2020
Data Scientist
Orchard Health, Boston
- Built the risk models for the care team, used for 40,000 patients each month.
Projects
- tinyserve
A small inference server for language models, written in Rust. It runs one model on one GPU with continuous batching.
Rust, CUDA, gRPC - evalcards
Short, repeatable evaluation reports for model releases, as static web pages.
Python, DuckDB - Batch scheduler notes
A written guide to batch scheduling for model serving, with measurements on three GPU types.
Writing
Code
@mira-sample on GitHub: 34 repositories, 2,690 stars, 812 followers.
- tinyserve
Small LLM inference server in Rust with continuous batching.
Rust, 1,840 stars, 96 forks - evalcards
Repeatable evaluation reports for model releases.
Python, 512 stars, 41 forks - kv-cache-viz
See how the KV cache fills during generation.
TypeScript, 187 stars, 12 forks - gpu-bench
Throughput and latency benchmarks for common GPUs.
Python, 96 stars, 9 forks - dotfilesShell, 31 stars, 2 forks
- triton-kernels
Fused attention and layer-norm kernels for study.
Python, 24 stars, 3 forks
Research
402 citations, h-index 5, 7 papers. Source: ORCID and Semantic Scholar.
- 2024Speculative Batching: Lower Latency for Shared Inference ServersMLSys214citations
- 2025When Do Evaluation Sets Go Stale? A Study of 40 Model ReleasesNeurIPS Datasets and Benchmarks88citations
- 2022Grasp Detection with Few Labels in Warehouse ScenesICRA61citations
- 2020Calibrated Risk Scores for Care ManagementMachine Learning for Health23citations
- 2026Draft Models That Know When to StoparXiv3citations
Education
- University of MichiganMS, Computer Science2016 to 2018
- University of LisbonBS, Mathematics2012 to 2016