Skip to content
View tao-hpu's full-sized avatar

Block or report tao-hpu

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 donโ€™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 userโ€™s behavior. Learn more about reporting abuse.

Report abuse
tao-hpu/README.md

Hey, I'm Tao An ๐Ÿ‘‹

Founder & CEO @ FIM Labs ยท MS AI ยท Enterprise AI Agents & Document Intelligence ยท China ร— Global


Building cognitive architectures that give LLMs better memory than mine ๐Ÿง 


๐Ÿ”ฌ About Me

class TaoAn:
    def __init__(self):
        self.location = "Singapore ร— Beijing"
        self.company = "FIM Labs"
        self.role = "Founder & CEO"
        self.education = "MS in Artificial Intelligence, Hawaii Pacific University (2026)"
        self.interests = ["RAG", "LLM Memory Architectures", "Knowledge Graphs", "Agentic Workflows"]
        self.current_focus = "Enterprise AI agents & document intelligence across the China ร— Global boundary"
        self.belief = "Most 'AI agents' should be workflows โ€” fix the plumbing before buying a bigger pump."

    def say_hi(self):
        print("Thanks for dropping by! Let's build something cool together.")

me = TaoAn()
me.say_hi()

๐Ÿ› ๏ธ Tech Stack

Python PyTorch LangChain OpenAI

TypeScript Next.js Docker Linux


๐ŸŒฑ Selected contributions

Product

  • fim-ai/fim-one โ€” open-source agent platform for Global ร— China enterprises (self-hosted, any LLM). Day-job / company product I build and ship.

Upstream (external)

  • openai/openai-agents-python stars โ€” two bugs in how a paused run is restored from serialized state:
    • #5142 (merged, ad93f54): a nested agent-as-tool run serializes its agent references relative to the tool's own agent, but resumption resolved them against the parent, so with two same-named agents a pending approval silently never applied. Root cause, 7-line fix, 218 lines of regression tests.
    • #3749: a pending nested approval could bind to the wrong tool call once an earlier entry was filtered out. The maintainer re-landed it at source level as #3753 with Co-authored-by: Tao An.
  • eigenpal/docx-editor stars โ€” CJK typography in the OOXML layout engine, ~2k lines merged: resolved the eastAsia font slot so CJK runs measure and paint in their own face (#539), and added kinsoku line breaking between ideographs (#540).
  • acmesh-official/acme.sh stars โ€” #7278 (merged, 6467ca9): _setopt() and _clear_conf() redirected new content straight into the conf file, so a write that failed on a full disk truncated it to 0 bytes while _save_conf still returned 0, and the certificate could no longer renew (#7247). Every conf write now goes through a temp copy in the same directory that is verified and then renamed over the original; symlinked or bind-mounted confs keep the in-place write. Checked under dash, bash, zsh and busybox ash, and on an ext4 loop mount filled to 0 free blocks.
  • docling-project/docling stars โ€” #4336 (merged, 2682193): the DOCX backend treated any list level with an East Asian w:numFmt as a bullet list, so ็ฌฌ%1ๆก-style regulation numbering was dropped on conversion (#4335). Added nine formats (chineseCounting, chineseLegalSimplified, japaneseCounting, decimalEnclosedCircle and others) following ECMA-376 ยง17.18.59, with the cases the standard leaves open settled against Microsoft Word's own output: the converters match 525 Word-rendered markers except one documented glyph choice, backed by 176 parametrized boundary tests and an end-to-end DOCX test.
  • docling-project/docling-core โ€” the follow-up the docling reviewers asked for: #799 (merged, 25ade75): the Markdown serializer's AUTO marker mode kept an original list marker only if it matched [a-zA-Z0-9], so the CJK markers docling now emits (็ฌฌไนๆก, ๏ผˆไธ€๏ผ‰, โ‘ ) were replaced by - (#798). The check now accepts any Unicode letter or digit, while bullet glyphs are still dropped; full suite 768 passed with ground-truth files unchanged.

Original research / tools

  • cognitive-workspace โ€” active memory / functional infinite-context architecture for LLMs.
  • nano-spec โ€” lightweight task-spec format for AI-assisted development.
  • cog-canvas โ€” training-free long-term memory for LLM conversations.

๐Ÿ Contribution Graph

Snake animation

๐Ÿ“Š GitHub Stats


๐Ÿ“ˆ Activity Graph


๐Ÿ”ฅ Coding Heatmap

coding heatmap

๐Ÿ“– Selected Research

For my latest papers and research, visit my website ๐Ÿ‘‰ tao-hpu.github.io

Website


Profile Views

Pinned Loading

  1. fim-ai/fim-one fim-ai/fim-one Public

    Open-source agent platform for Global ร— China enterprises โ€” wire every system through one agent core. Self-hosted, any LLM.

    Python 1.6k 178

  2. ccmap ccmap Public

    Coding heatmap for Claude Code + Codex โ€” scan local usage, render a GitHub-style heatmap badge & shareable HTML report

    TypeScript 3 1

  3. cognitive-workspace cognitive-workspace Public

    ๐Ÿง  Active memory management system enabling functional infinite context for LLMs through cognitive workspace architecture

    TeX 105 11

  4. nano-spec nano-spec Public

    Spec-driven thinking, nano-sized docs. Lightweight task specification for AI-assisted development.

    49 5

  5. cog-canvas cog-canvas Public

    CogCanvas is a training-free framework for maintaining long-term memory in LLM conversations.

    Python 9 1

  6. llm-from-scratch llm-from-scratch Public

    ไปŽ 0 ๅˆฐ 1 ๆ‰‹ๆ“ๅคงๆจกๅž‹:้€่กŒไธญๆ–‡่ฎฒ่งฃ,ไปŽๅญ—็ฌฆ็บง bigram ไธ€่ทฏๅค็Žฐ GPT-2 124M(ๅซไบคไบ’ๅผๅฏ่ง†ๅŒ–)ใ€‚A from-scratch LLM course in Chinese โ€” build a Transformer by hand, then reproduce GPT-2 124M.

    HTML 4