Skip to content
View mynotwo's full-sized avatar
🥰
🥰

Block or report mynotwo

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
mynotwo/README.md

Hi, I’m Audrey — compression, inference systems and agent reliability

I publish as Yu Mao. I'm a Research Scientist at ByteDance working on AI infrastructure: LLM inference and reproducibility, compression and representation, model capability, and agent reliability. I received my Ph.D. in Computer Science from City University of Hong Kong and am based in San Jose.

Homepage · Publications · Google Scholar · LinkedIn

Research projects

Project Status Question Start with
KVBytePack Reproducible artifact When do smaller KV records pay for their codec cost? Public records, CPU cost and memory measurements
BitExact LLM Research prototype What stays identical when inference execution changes? Reproduction guide · Scoped vLLM plugin evidence
AIBrix Hybrid KV Cache Experimental What state must a hybrid prefix cache preserve? CPU storage/trace example · RFC · Open draft PR #3221
Random-Reward RL Probe Paper artifact What can training reach without correctness feedback? Paper result CSVs and GPU-free figure reproduction

These projects expose their tradeoffs and evidence: byte-preserving codecs, explicit inference-identity contracts, independently testable storage behavior, and reported training outcomes. The open draft AIBrix PR proposes only the byte-store contract, excludes the hybrid serving connector and codecs, and has not been merged. The hybrid integration's model equivalence is still unresolved; BitExact's current H800 plugin runs are scoped to the recorded models, versions and configurations; its CPU checks and device reports do not certify every device.

Earlier compression work includes TRACE (WWW 2022), with follow-ups at ACM MM 2022 and DAC 2023.

Everyday tools

I'm also building Hushword: speak in the app you're using, then return to writing. It combines local Whisper with Apple Speech and offers an optional Agent rewrite after dictation.

Build and try it on an Apple Silicon Mac · Where audio and text are processed

📝 Selected publications

Full list on my homepage and Google Scholar.

🛠️ Tools I use

Python · PyTorch · Triton · CUDA · vLLM · Go · eBPF

Approximate profile views, counted from image requests

Pinned Loading

  1. kvbytepack kvbytepack Public

    Measure when byte-exact KV-state compression pays off: public records, complete-record CPU and memory costs, and GPU-free reproduction.

    Python 2

  2. bitexact-llm bitexact-llm Public

    Study bitwise reproducibility with experimental integer Llama inference, weight-free CPU checks and scoped device reports.

    Python 1

  3. aibrix-hybrid-kvcache aibrix-hybrid-kvcache Public

    Experimental hybrid KV offloading with variable-length records, bounded MEMORY storage and a verifiable CPU byte path. Model equivalence remains unresolved.

    Go 1

  4. rl-random-prob rl-random-prob Public

    Random-reward RL as a probe of checkpoint reachability: paper code, reported result CSVs and GPU-free figure reproduction.

    Python

  5. hushword hushword Public

    macOS menu bar dictation with local Whisper, Apple Speech and optional Agent rewrites. Apple Silicon; build from source.

    Swift

  6. A-Fast-Transformer-based-General-Purpose-LosslessCompressor A-Fast-Transformer-based-General-Purpose-LosslessCompressor Public

    TRACE: Transformer-based general-purpose lossless compression (WWW 2022). Research code, dataset link and compression example.

    Python 29 6