AI research knowledge base — papers, concepts & intuitions Designed by Artificial Intelligence Research Center, Hon Hai Research Institute
🏛️ Institute: https://hhri.foxconn.com/en 🌐 Live Demo: https://MuhammadSaqlainAslam.github.io/my-llm-wiki
This is a personal AI/ML research knowledge base built from academic papers, concept glossaries, and cross-linked notes. Every paper is distilled into a structured Obsidian note with YAML frontmatter (title, authors, year, arXiv ID, citation count, tags, TL;DR), a full intuition-first explanation, and [[wikilinks]] connecting it to every related concept. The result is a single navigable graph where clicking any node — whether a paper or a sub-concept — reveals exactly how it fits into the broader ecosystem.
The writing philosophy is borrowed from Andrej Karpathy: lead with the concrete problem and why it matters, introduce the math only after the intuition is clear, and always give real numbers. Every note leads with intuition, backs claims with concrete numbers, and skips filler — the same standard used in the best ML writing.
What makes this wiki distinct is a full citation intelligence layer on top of the content. Every core paper carries its real citation count from Semantic Scholar, a cited_by_details list of the ten most-cited downstream works, and arXiv IDs throughout. The live web demo exposes all of this through a Citation Explorer, an interactive D3 knowledge graph sized by citation count, and a full-text search interface — all from a single static site deployed automatically via GitHub Actions to GitHub Pages.
Citation counts sourced from Semantic Scholar. Last verified: June 2026. Auto-updated monthly via
python3 agent.py update-citations.
The most referenced note in the wiki is "Attention Is All You Need" with 46 backlinks — reflecting its foundational role across all modern LLM research.
- Full-text search, multi-tag filtering, and sort controls across all notes
- Citation Explorer — master-detail panel with ranked citation bars and top-10 citing papers
- Interactive D3.js knowledge graph sized and colored by citation count and research theme
- Theme Navigation — notes grouped by research area
- Three appearance modes: System / Dark / Light with live theme switching
- Reading progress tracker with per-note status badges (Unread / Reading / Done)
- Timeline view — papers arranged chronologically 2017–2026
- Knowledge graph node color coding by research theme with color and size legends
- Full-text search across all notes by title, tag, or keyword
- Multi-tag AND filtering — click multiple tags to narrow results simultaneously
- Note cards showing title, year badge, citation count badge (🔖 N citations), tags, and 3-line TL;DR
- Sort by title A–Z, year (newest first), or citation count descending
- Side panel opens the full note — rendered Markdown with KaTeX math, live
[[wikilink]]resolution, and clickable cross-references
A master-detail two-panel layout — click the Citation Explorer tab to open it.
Left panel — ranked list of every paper with citation_count > 0, sorted descending:
- Each row:
#rank · title · year · proportional citation bar · count - Bar width proportional to citations (
citation_count / 100,000), minimum 4 px so even small counts are visible - Bar color by impact tier:
- 🟣 Purple → 10,000+ citations
- 🩵 Teal → 1,000–9,999 citations
- 🔵 Blue → 100–999 citations
- ⬜ Gray → under 100 citations
- Filter box at top to search the ranked list in real time
Right panel — default state (nothing clicked): summary dashboard showing:
- Total papers in wiki
- Total citations tracked
- Most cited paper
- Fastest growing paper (Mamba: 7,268 citations since 2023)
- Citation cluster pills: Foundations · Efficiency · Inference · Hardware
Right panel — paper selected:
- Title, authors, year, and arXiv badge with external link
- Large citation count number + full-width color bar
- What this paper introduced — the note's TL;DR
- Top 10 papers that cite this work — ranked table with: title (clickable), year, ~citations, theme badge, arXiv ↗ link
- Clicking a citing paper's title: loads its detail view if it exists in the wiki, otherwise opens
arxiv.org/abs/[id]in a new tab
- Clicking a citing paper's title: loads its detail view if it exists in the wiki, otherwise opens
- Related papers in this wiki — all resolved
[[wikilinks]]from the note shown as clickable pills; clicking one switches to the All tab and opens that note's detail panel - ← All Papers back button returns to the summary dashboard
- Interactive D3.js v7 force-directed graph (
graph.html) - Every node = one wiki note, sized logarithmically by
citation_count; nodes without citation data sized by incoming link count - Every edge = a
[[wikilink]]connection between notes - Color-coded by theme tag (foundations, efficiency, hardware, inference-optimization, scaling, synthesis, meta)
- Click any node → opens full note detail in the side panel with two action buttons:
- 📄 Read on arXiv — opens the paper on arxiv.org (shown when arXiv ID is available)
- 🔗 Open in wiki — opens the full note in the browse view
- Hover a node → highlights direct neighbors, dims everything else; tooltip shows title, year, citations, TL;DR, and arXiv ID if available
- Filter by tag — dims non-matching nodes
- Legend: color = theme, node size = citation count (where available, else link count)
- Zoom, pan, drag nodes freely
- Node color coding by research theme:
- Purple → SSM / Mamba / sequence modeling papers
- Teal → concept glossary stubs
- Amber → inference optimization papers
- Blue → hardware & systems papers
- Green → foundational papers
- Red → benchmarks & evaluation
- Color legend displayed in the top-left corner of the graph
- Node size legend displayed in the bottom-left:
- Small = under 100 citations (sized by incoming link count)
- Medium = 100–1,000 citations
- Large = 1,000–10,000 citations
- Extra large = 10,000+ citations (e.g. Attention Is All You Need)
- Theme colors update live when appearance mode is switched — no page reload
Notes grouped by theme in the By Theme tab:
| Theme | Contents |
|---|---|
| Foundations | Self-attention, RLHF, open models |
| State Space Models | SSMs, linear RNNs, LSTM variants |
| Mixture-of-Experts | Sparse routing, load balancing |
| Inference Optimization | Speculative decoding, KV cache, draft models |
| Hardware & Systems | GPUs, TPUs, FPGAs, ASICs, LPUs, IO-aware kernels |
| Attention & KV | Attention variants, KV cache techniques |
| Training & RL | RLHF, RLVR, GRPO, optimizers |
| Benchmarks | Frontier model comparisons |
| Concepts & Glossary | 100+ sub-concept stub notes |
- Three-way toggle in the nav bar: System / Dark / Light
- System: automatically follows your OS preference (
prefers-color-scheme) - Dark: deep background (
#0d1117), glowing colored nodes, high-contrast text - Light: clean white background, rich color nodes, dark labels
- Preference saved between sessions via
localStoragekeyappearance - All components respond: cards, graph nodes, citation bars, side panel, legends
- Track your reading status per note directly in the browser — no account needed
- Three states per note: Unread / Reading / Done
- Click the status badge on any note card to cycle through states
- Progress bar at the top of the browse view: "X of Y notes done · M reading"
- Filter the grid by reading status: All / Unread / Reading / Done
- Progress filter applies across All, By Theme, and Timeline tabs
- State persisted in
localStorage— survives page reloads and browser restarts
- Horizontal scrollable timeline spanning 2017 to 2026
- Papers arranged in columns by publication year — one column per year
- Each card shows: title, up to 3 tags, and theme color stripe
- Click any card to open the full note detail panel
- Color-coded by research theme — same scheme as the knowledge graph
- Concept stubs and reference notes without a year in a dedicated Reference column
- Reveals the field's evolution at a glance:
- 2017 — Transformer (Attention Is All You Need)
- 2022 — FlashAttention, S4 Structured State Spaces
- 2023 — Mamba, RWKV, RetNet, LLaMA 2, FlashAttention-2
- 2024 — xLSTM, Griffin, Mamba-2, Medusa, EAGLE
- 2025–2026 — Speculative Decoding, KV Cache Optimization, DeepSeek-V4, Hardware Acceleration
- KaTeX renders LaTeX inline (
$O(n^2)$) and display ($$\text{Attention}(Q,K,V) = \text{softmax}\!\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$) - Math is protected during Markdown preprocessing so the renderer never mangles LaTeX delimiters
- Supports
$...$,$$...$$,\(...\),\[...\]delimiters
- Indexes complete note body content — not just title and tags
- Results ranked by relevance: title match → tag match → tldr → body content
- Press
/anywhere on page to focus search bar - Result count shown: "X notes found"
Ctrl+F style match navigation:
- Opening a note from search auto-scrolls to the first match instantly
- Match navigation bar appears at top of note: 🔍 "search term" 1 of 6 ‹ ›
- ‹ › buttons navigate between every occurrence
- Current match = bright purple highlight
- All other matches = light purple highlight
- Keyboard shortcuts: Enter or ↓ → next match · Shift+Enter or ↑ → previous match
- Navigation bar hides automatically when note is closed or search is cleared
- Case insensitive — finds "Attention", "attention", "ATTENTION" equally
- Single-match notes disable the › button automatically
Example workflow:
- Type
"exponential gating"in search bar - xLSTM note appears in results (term is in the body, not the title)
- Click xLSTM → note opens, auto-scrolls to first match, shows "1 of 3 ‹ ›"
- Press › to jump through all 3 occurrences
- Clear search → highlights disappear
- Click ⊕ Compare on any note card to select it for comparison
- Select 2 or 3 papers — a comparison tray slides up from the bottom of the page showing selected titles as pills
- Click "Compare (X)" to open a full side-by-side table view
- Toggle between Summary (quick 5-row) and Technical (10-row default) views
- Technical mode shows 10 rows extracted from each note:
- Core Innovation, Architecture Type, Sequence Complexity, Key Formula / Mechanism, Best Benchmark Result, Hardware Requirement, Limitations, vs Transformer, Year & Citations, Reading Status
- Key differences: auto-detected — a banner highlights which rows actually differ between papers
- Row labels color-coded: purple when papers differ (key differences), gray when they match
- Column top borders color-coded: first paper = purple, second = teal, third = amber
- Alternating row backgrounds for readability
- Maximum 3 papers at once; selecting a 4th bumps the oldest
- Press Escape or click ✕ Close to dismiss
- Every note tracks which other notes reference it — the inverse of the
[[wikilink]]graph - "Referenced by X notes" section appears at the bottom of every open note detail panel (only shown when X > 0)
- Clicking any backlink pill opens that note in the same panel
- First 8 backlinks shown; click "+X more" to expand the full list
- ← X backlink count badge on every note card
- Example: "Attention Is All You Need" shows ← 46, meaning 46 other notes reference it — the most central paper in the wiki
| Paper | Year | ~Citations | arXiv |
|---|---|---|---|
| Attention Is All You Need | 2017 | ~178,574 | arXiv:1706.03762 |
| LLaMA 2 | 2023 | ~17,078 | arXiv:2307.09288 |
| Paper | Year | ~Citations | arXiv |
|---|---|---|---|
| S4: Structured State Spaces | 2022 | ~3,636 | arXiv:2111.00396 |
| RWKV | 2023 | ~1,063 | arXiv:2305.13048 |
| RetNet | 2023 | ~670 | arXiv:2307.08621 |
| Mamba | 2023 | ~7,268 | arXiv:2312.00752 |
| Griffin | 2024 | ~239 | arXiv:2402.19427 |
| Transformers are SSMs (Mamba-2) | 2024 | ~1,526 | arXiv:2405.21060 |
| xLSTM | 2024 | ~595 | arXiv:2405.04517 |
| Paper | Year | ~Citations | arXiv |
|---|---|---|---|
| Switch Transformers / Mixtral | 2022 | — | — |
| Paper | Year | ~Citations | arXiv |
|---|---|---|---|
| FlashAttention | 2022 | ~4,404 | arXiv:2205.14135 |
| FlashAttention-2 | 2023 | ~2,769 | arXiv:2307.08691 |
| Hardware Acceleration for Neural Networks | 2025 | ~3 | arXiv:2512.23914 |
| Paper | Year | ~Citations | arXiv |
|---|---|---|---|
| Medusa | 2024 | ~728 | arXiv:2401.10774 |
| EAGLE | 2024 | ~474 | arXiv:2401.15077 |
| Speculative Decoding | 2025 | ~6 | arXiv:2601.11580 |
| KV Cache Optimization | 2026 | ~1 | arXiv:2603.20397 |
| Paper | Year | ~Citations | arXiv |
|---|---|---|---|
| Nemotron-3 | 2025 | — | — |
| DeepSeek-V4 | 2026 | — | — |
| LLM Benchmarks | 2025 | — | — |
Over 100 concept stub notes cover every sub-component referenced across the paper articles, organised into six clusters: Attention & Transformers (multi-head attention, GQA, MQA, RoPE, XPOS, online softmax, tiling, recomputation, IO-awareness, warp, occupancy, thread block, linear attention, local attention, sliding window attention, attention sinks, CSA, HCA); State Space Models (State Space Model, HiPPO, HiPPO matrix, Cauchy kernel, discretization, convolutional view, Selective SSM, semiseparable matrix, multi-head SSM, retention mechanism, chunkwise recurrent, RG-LRU, LRU, exponential gating, matrix memory, covariance update rule, sLSTM, mLSTM, Long Range Arena); Inference Optimization (KV Cache, paged attention, KV offloading, prefill, decode phase, cache eviction, cache compression, hybrid memory, eviction policy, H2O eviction, draft model, verification step, lookahead decoding, tree attention, Medusa heads, Medusa-1, Medusa-2, lossless speedup, feature-level drafting, one-step token advance, multi-token prediction); Hardware & Systems (GPU, TPU, NPU, FPGA, ASIC, LPU, systolic array, tensor cores, HBM, SRAM, memory hierarchy, kernel fusion, operator fusion, hardware-aware scan, NVFP4, pallas kernel, in-memory computing, near-memory computing, neuromorphic computing); Training & Optimization (RLHF, RLVR, GRPO, ghost attention, load balancing loss, on-policy distillation, Muon optimizer, tensor parallelism, sequence parallelism); Architecture Components (LatentMoE, Manifold-Constrained Hyper-Connections, selective state spaces, LLM evaluation, LLM benchmarks). Every concept links back to the paper where it originates, and every [[wikilink]] in a paper note resolves to a concept page — 0 orphan links.
The wiki includes a full citation intelligence layer: wiki/Citation Map.md tracks total Semantic Scholar citation counts for the six core papers plus their top-10 highest-impact downstream works. Every core paper's YAML frontmatter carries citation_count (integer), arxiv (arXiv ID string), cited_by_top (name list), and cited_by_details (list of dicts: title, year, citations, theme, arxiv) — all emitted by build.py into notes.json and consumed by the Citation Explorer in the web demo. The wiki/🗺️ Dashboard.md Obsidian dashboard includes a live Citation Leaderboard Dataview query. Papers with full cited_by_details: Attention Is All You Need, Mamba, xLSTM, Speculative Decoding, KV Cache Optimization, Hardware Acceleration for Neural Networks.
my-wiki/
├── wiki/ # Markdown notes — grows automatically (Obsidian vault)
│ ├── Home.md # Entry point with concept map + reading paths
│ ├── 000 Index.md # Full index: themes, glossary, narrative, citation table
│ ├── Citation Map.md # Citation scores + top-10 citing papers per core paper
│ ├── 🗺️ Dashboard.md # Live Dataview queries (Obsidian only)
│ ├── Mamba.md # Paper notes (20 total)
│ ├── Transformer.md
│ ├── ...
│ ├── KV Cache.md # Concept stub notes (100+ total)
│ ├── HBM.md
│ └── ...
│
│ 💡 Grows automatically — run `python3 agent.py topic "..."` to add papers.
│ Note count updates live on the badge at the top of this README.
│
├── docs/ # Static web demo (auto-deployed)
│ ├── index.html # Browse, search, Citation Explorer
│ ├── graph.html # D3 knowledge graph
│ └── notes.json # All notes as structured JSON
├── raw/ # Source PDF papers
├── build.py # wiki/*.md → docs/notes.json + HTML
├── build_wiki.py # PDF → wiki note (Claude API)
├── agent.py # Agentic ingestion system
├── agent_log.json # Log of all agent runs
├── setup_cron.sh # Optional cron scheduler
├── cron.log # Cron output log (if scheduled)
├── .github/workflows/deploy.yml # Auto-deploy to GitHub Pages on push
└── README.md
PDF papers in raw/
↓
build_wiki.py — Claude API extracts & summarises
↓
wiki/*.md — structured notes with YAML frontmatter
↓
build.py — converts Markdown + math + wikilinks → JSON + HTML
↓
docs/notes.json + docs/index.html + docs/graph.html
↓
GitHub Actions — auto-deploys on every push to main
↓
https://MuhammadSaqlainAslam.github.io/my-llm-wiki
build.py handles: YAML frontmatter parsing, math protection (swap LaTeX for placeholders before Markdown rendering → restore raw LaTeX for KaTeX), Obsidian callout → HTML div conversion, ![[embed]] → clickable link conversion, Dataview/Mermaid block stubbing, and emission of id, title, authors, tags, year, tldr, aliases, links, citation_count, arxiv, cited_by_details, and rendered html per note.
The wiki has an autonomous research agent (agent.py) that automatically finds, evaluates, and adds new content from multiple sources using Claude as the decision-making brain.
1. Topic search — add content about a specific topic:
python3 agent.py topic "DeepSeek R2 architecture"
python3 agent.py topic "LLM inference optimization 2025"
python3 agent.py topic "Mamba 2 SSM improvements"2. Citation tracking — find high-impact papers citing existing wiki papers:
python3 agent.py citationsQueries Semantic Scholar API for papers with 100+ citations that cite existing wiki papers. Automatically downloads any high-impact papers not yet in the wiki.
3. Daily monitor — broad sweep of new content:
python3 agent.py dailySearches arXiv (cs.LG, cs.CL, cs.AI), checks GitHub repos for updates, fetches blog posts. Run this once a week for a broad update.
4. Update citation counts — refresh from Semantic Scholar:
python3 agent.py update-citationsQueries the Semantic Scholar batch API for every note that has an arXiv ID in its frontmatter, updates citation_count in the YAML, rebuilds the web demo, and commits + pushes automatically. This is a deterministic API call (no LLM), so it's fast and cheap. Run monthly to keep counts current.
| Command | Frequency | Purpose |
|---|---|---|
python3 agent.py update-citations |
Monthly | Keep citation counts current |
python3 agent.py citations |
Monthly | Find new high-impact citing papers |
python3 agent.py daily |
Weekly | Broad new-paper sweep |
python3 agent.py topic "..." |
On demand | Add specific topic papers |
| Source | What it fetches | API |
|---|---|---|
| arXiv | New papers by keyword/topic | Free |
| Semantic Scholar | Citation tracking | Free |
| GitHub | READMEs, notebooks, releases | Free |
| Blogs | Karpathy, Lilian Weng, HuggingFace, Distill | Web fetch |
| YouTube | Lecture transcripts (yt-dlp) | Free |
The agent is intentionally triggered manually rather than running on a cron schedule:
- Full control over what enters the knowledge base
- No unnecessary server load on shared infrastructure
- Every update is intentional and visible
- API credits used only when needed
Every agent run is logged to agent_log.json:
# See what was added in the last run
cat agent_log.json
# See cron output if scheduled
cat cron.logClaude evaluates every candidate paper or article against these wiki themes before adding: transformer, attention, SSM, Mamba, LSTM, speculative decoding, KV cache, MoE, FlashAttention, hardware acceleration, LLM inference, scaling, RLHF, tokenization, language models
Only content scoring 7/10+ relevance is added. Everything else is logged as skipped.
# Weekly: broad new paper sweep
python3 agent.py daily
# On demand: specific topic
python3 agent.py topic "your research interest"
# Monthly: citation updates
python3 agent.py citations- Drop the PDF into
raw/ python build_wiki.py— Claude API extracts and writeswiki/<Title>.mdpython build.py— regeneratesdocs/notes.jsongit add . && git commit -m "Add <Paper>" && git push- GitHub Actions deploys to GitHub Pages in ~2 minutes
Copy .env.example and fill in your own values — never commit real keys:
cp .env.example .env
# then edit .env, or export the variables in your shell# Option 1 — Anthropic public API
export ANTHROPIC_API_KEY="sk-ant-..."
# Option 2 — Google Vertex AI (enterprise)
export ANTHROPIC_VERTEX_PROJECT_ID="your-gcp-project-id"
export VERTEX_REGION_CLAUDE_4_6_SONNET="europe-west1"
# Wiki location (optional — defaults to the repo directory)
export WIKI_VAULT="/path/to/your/wiki"pip install anthropic pymupdf markdown pyyaml
python build_wiki.py # process new PDFs → wiki/*.md
python build.py # regenerate docs/ from wiki/
open docs/index.html # preview in browserFor the knowledge graph, serve locally to avoid CORS issues:
python -m http.server 8000
# then open http://localhost:8000/docs/graph.html| Component | Tool |
|---|---|
| Note generation | Claude API (claude-sonnet-4-6) |
| PDF extraction | PyMuPDF |
| Markdown → HTML | Python markdown library |
| Math rendering | KaTeX (client-side) |
| Knowledge graph | D3.js v7 (force simulation) |
| Static site | Vanilla HTML / CSS / JS — zero build step, zero dependencies |
| Hosting | GitHub Pages |
| Local knowledge base | Obsidian (Dataview plugin) |
| Version control | Git + GitHub |
| Auto-deploy | GitHub Actions |
| Reading tracker | localStorage API |
| Timeline view | Vanilla JS + CSS flex |
| Appearance modes | CSS prefers-color-scheme + JS |
| Research agent | Claude API tool use |
| arXiv API | urllib + XML parsing |
| Semantic Scholar | REST API |
| yt-dlp | Python library |
| GitHub API | REST API |
Designed by the Artificial Intelligence Research Center
Visitor counter: Live country-based visitor tracking displayed at the bottom of every page via Flag Counter — updates automatically on every visit, visible to all visitors.
Analytics dashboard: Privacy-friendly analytics via Cloudflare Web Analytics. No cookies. No personal data. GDPR compliant. View at: dash.cloudflare.com → Web Analytics
Muhammad Saqlain Aslam GitHub: https://github.com/MuhammadSaqlainAslam Repository: https://github.com/MuhammadSaqlainAslam/my-llm-wiki Web Demo: https://MuhammadSaqlainAslam.github.io/my-llm-wiki