Skip to content
View cassieliang6709's full-sized avatar
🌴
On vacation
🌴
On vacation

Highlights

  • Pro

Block or report cassieliang6709

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. Markdown is supported. 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
cassieliang6709/README.md

Hi, I'm Cassie.

M.S. in AI at Northeastern, SWE intern at Tiiny AI · a video diary you can make with friends · one memory shared by Claude Code and Codex · now learning AI infra and RL post-training · GRPO on GSM8K: 34.7% to 52.2% on one RTX 4090

AI master's student at Northeastern, SWE intern at Tiiny AI. I'm focusing on LLM inference and training systems — reading SGLang and nano-vllm, working through CS336, and running small, controlled experiments on one GPU. On the side I ship products: 1Day is on the App Store.

Portfolio · Notes · LinkedIn · Email

Open source

Project Contribution Status
SGLang Cache CPU image-preprocessing artifacts for MiniMax-M3 multimodal serving — #44181 🟡 Open
OpenLess Prefer the microphone's native format and reject empty recordings — #1135 ✅ Merged
Raycast-Easydict Send favorite words to Anki — #109 ✅ Merged
modular-diary 6 PRs: checkbox-list todos, drag-to-timeline, current-time marker, CI test-runner fix — #2–#7 ✅ Merged

LLM systems & research

GRPO on GSM8K — baseline → LoRA SFT → GRPO reproduction on a single RTX 4090: 34.7% → 52.2%. Code

Stop-token study — does a chat model learn to stop if its end token never appeared in pretraining? A controlled MiniMind pretraining experiment. Code

Agent honesty eval — does a coding agent admit a task is impossible? MBPP-derived tasks with contradictory tests, a tamper-proof grader, Qwen2.5-Coder 3B/7B results. Code

CorpCheck — SEC filing search that knows when not to answer: vector + full-text retrieval fused with RRF, revision-aware filtering, deterministic IR evaluation, refusal on weak evidence. Code · Demo

Products

1Day — a video diary you can make with friends. Follow filming prompts and the app assembles a short film entirely on-device (AVFoundation); shared stories sync through CloudKit. Built end to end. App Store · Website · Code

MindBridge — one memory shared by Claude Code and Codex. Turns agent logs into searchable memories in PostgreSQL / pgvector, served over MCP. Code · Demo

VisoCode — LangGraph agents that write, audit, run and repair Manim lessons in Docker sandboxes. AdventureX 2025 Best Technical Award. Code

OpenTranscript — a searchable transcript beside any YouTube / Bilibili video, exported as a Markdown note with clickable timestamps. Code · Download

Vance — an AI fitness coach that edits your workout mid-session through allowlisted tool calls. Code · Demo

Work

Tiiny AI — SWE intern, Jul 2026–present. Agent skills platform: catalog APIs, desktop install/update flow, and a Rust/Python distribution backend with versioning, validation, rollback and permission checks.

Northeastern — TA, CS6180 Foundations of Generative AI.

Deloitte — audit assistant, 2023–2024. Automated ledger reconciliation; encoded recurring review checks as configurable rules.

Stack

Python · TypeScript · Swift · Rust · SQL — PyTorch, SGLang, MLX · FastAPI, PostgreSQL / pgvector, Redis, Docker · RAG, MCP, LoRA / GRPO · SwiftUI, AVFoundation, React

Pinned Loading

  1. 1day-ios 1day-ios Public

    Native iOS app that turns daily moments into collaborative short films on-device.

    Swift 2

  2. fitness-coach-ios fitness-coach-ios Public

    AI 健身教练 — 陪练模块 (SwiftUI iOS, mock data)

    Swift 1

  3. open-transcript open-transcript Public

    Local-first video transcript workbench for YouTube and Bilibili.

    JavaScript 1 1