title	LearnLens
emoji	🔍
colorFrom	blue
colorTo	purple
sdk	gradio
sdk_version	5.0
app_file	learnlens/app.py
pinned	false
license	apache-2.0
hardware	a10g
short_description	AI learning mentor that tells you what to focus on
tags	learnlens learning-mentor small-models minicpm nemotron knowledge-distillation

LearnLens 🔍

AI learning mentor powered by 3 small models.

LearnLens is not a bookmark manager. It is an opinionated coach that tells you what to focus on, what you're forgetting, and what mistakes to watch for.

---

Why LearnLens?

Every day, ML engineers and researchers consume dozens of articles, papers, tweets, and videos. By the next morning, 90% is forgotten.

• ChatGPT/Claude can't access your browsing history or remember your mistakes
• Obsidian/Notion are passive libraries -- they store, but never judge or prioritize

LearnLens runs locally, reads your content, and actively tells you when you're wasting time.

---

Architecture

User Input (URLs / Bookmarks / Chrome History)
         |
    Ingestion (trafilatura / BS4)
         |
    +------------------+
    |   SQLite DB      |
    +------------------+
         |
    Model 1: Connector (nomic-embed-text-v1.5, 137M)
         |  -> embeddings, similarity, forgotten items
         |
    Model 2: Prioritizer (MiniCPM5-1B + LoRA-scorer, ~1B)
         |  -> structured JSON scoring 1-10 vs goals
         |
    Model 3: Mentor (MiniCPM5-1B + LoRA-mentor, ~1B)
         |  -> opinionated markdown daily briefing
         |
    Gradio UI (5 tabs)

Key innovation: Models 2 and 3 share the same MiniCPM5-1B base weights (~2GB) and swap LoRA adapters at runtime. Total parameters: ~2.2B (Tiny Titan track eligible).

---

Features

• Daily Briefing — opinionated markdown report on what to read today
• Priority Queue — all content scored 1-10 against your goals
• Forgotten Items — high-relevance content you haven't touched in 7+ days
• Knowledge Map — UMAP visualization of your content embeddings
• Mistake Tracker — track patterns you want to avoid repeating

---

Quick Start

# Install dependencies
uv sync

# Run locally
uv run learnlens

# Seed demo data
uv run python scripts/seed_demo_data.py

---

Tech Stack

Component	Technology
UI	Gradio 5.0+
Embedding	nomic-embed-text-v1.5 (137M)
Prioritizer	MiniCPM5-1B + LoRA-scorer (~1B)
Mentor	MiniCPM5-1B + LoRA-mentor (~1B)
Shared Base	MiniCPM5-1B loaded once, 2 adapters swapped
Storage	SQLite (WAL mode)
Training	Modal + TRL + NVIDIA NIM

---

Training

# Generate distillation data (requires NVIDIA_NIM_API_KEY)
uv run python training/generate_data.py
uv run python training/generate_mentor_data.py

# Fine-tune both adapters on Modal
modal run training/train_prioritizer.py
modal run training/train_mentor.py

---

Sponsor Alignment

Sponsor	How We Qualify
OpenBMB	MiniCPM5-1B as shared base for both prioritizer and mentor
Tiny Titan	~2.2B total parameters (137M + 1B + 1B)
Backyard AI	Local-first, privacy-preserving
Llama Champion	GGUF export for llama.cpp
NVIDIA	49B -> 1B knowledge distillation via Nemotron Super
HuggingFace	Gradio, HF Spaces, PEFT, TRL

---

License

Apache 2.0