// swappable
Model ≠ lock-in
Only data artifacts are learned — weights, prompts, sources. Never the application code. The config survives any model swap.
Local-first · Rust · local LLMs
A morning briefing that
improves itself every day.
ibrief builds a curated briefing overnight, tailored to one specific person — then measures what was actually useful in order to adapt. Five minutes of signal instead of hours of noise. Fully local on an Apple Silicon machine.
- grow
- stay current
- decide well
- better conversations
Design philosophy
The memory is the product — not the model.
ibrief follows an "LLM-OS" stance: the LLM is a swappable compute core, the value lives in persistent, versioned state. Self-improvement stays on a leash — propose, verify against real feedback, safeguard, adopt, roll back.
// autonomy slider
On a leash
Every proposed change passes a safety gate before becoming the default. Verification is cheaper than generation — propose a lot, adopt little (only what's checked).
// anti-bubble
Engagement isn't the goal
A naive engagement optimizer makes you more comfortable but dumber. ibrief deliberately sacrifices short-term score to preserve diversity and anti-bubble invariants.
// gated autoresearch
Depth, controlled
Multi-step research (Karpathy-style) is powerful but expensive and hallucination-prone — so it is a budgeted, citation-required module, not a free-running agent.
The self-improvement circle
Deliver → Feedback → Evaluate → Learn → Safeguard.
The same loop runs every night. What was useful gets reinforced; what wasn't fades — within hard safety bounds.
01
Deliver
build & push the briefing
02
Feedback
👍/👎 via Telegram buttons
03
Eval
behavior + judge + structure
04
Learn
weights · prompts · sources
05
Gate
check · adopt · rollback
↺ The behavioral score is ground truth; the LLM judge mainly helps in cold-start and is calibrated against real feedback.
Three axes, each safeguarded
Self-improvement isn't one feature — it's three.
AxisLeverSafeguard
Weights
sources & topics
sources & topics
Multi-armed bandit (Thompson / Beta sampling) with an exploration floor [0.2, 2.0]
Safety gate: bounds + source diversity (max 50%), versioned, rollback
Prompts
e.g. TL;DR synthesis
e.g. TL;DR synthesis
An optimizer LLM generates variants; a shadow test scores both on the same input
Adopted only on a clear judge margin; prompt versioning + experiment log
Sources
feed registry
feed registry
Quality scoring from feedback + selection frequency; weak ones retired
Drift watchdog (HHI): on narrowing, pruning is suspended — breadth over score
The nightly pipeline
From raw feed to curated briefing.
Ingest→Dedup→Enrich→Score (weights)→Curate→TL;DR (learned prompt)→Render→Persist→Push
The bulk tier (summarizing, tagging) runs on a fast small model, the synthesis tier (curation, judge) on a 70B-class model — both local. The entire learning cycle runs overnight essentially for free; a frontier model is pulled in only optionally for judge calibration through an existing subscription (claude -p).
Anatomy of a briefing
Three sections with clear intent.
The last two cannot be optimized away by personalization. The learning loop cannot optimize you into a comfortable bubble.
Highlights
The most important stories of the day — personalized through the learned weights. Plus "The 3 things today" as an executive summary.
Counterpoint fixed
A seriously argued, steel-manned counter-argument to your own position — so you know the other side in conversation instead of arguing inside the bubble.
Wildcard fixed
A deliberately surprising, real article outside your usual interests — preferably from a source that didn't make the main selection.
Roadmap — fully implemented
Six milestones, each useful on its own.
M1 ✓
Static briefing
Ingest → Enrich → Curate → Render against local Ollama.
M2 ✓
Persistence & feedback
SQLite (sqlx), dedup, briefing records; Telegram push with 👍/👎 buttons.
M3 ✓
Eval engine
Behavioral score + LLM judge against a versioned rubric + deterministic structural checks.
M4 ✓
Learning: weights
Thompson bandit over sources/topics, safety gate, config versioning & rollback.
M5 ✓
Learning: prompts
Optimizer LLM + shadow test; a temporal A/B decision primitive.
M6 ✓
Source evolution + AutoResearch
Scoring, pruning, drift watchdog; gated deep-research loop with citation verification.
Quick start
Local in two minutes.
Requirement: Ollama with one bulk and one synthesis model. On smaller hardware just put smaller models in config/profile.toml.
# pull models ollama pull qwen2.5:14b # bulk tier (enrich) ollama pull llama3.3:70b # synthesis tier (TL;DR, judge) # build a briefing cargo run -p ibrief-app # = brief # learn & optimize cargo run -p ibrief-app -- eval # score today's briefing cargo run -p ibrief-app -- learn # learn weights (Thompson + gate) cargo run -p ibrief-app -- optimize # improve the TL;DR prompt via shadow test cargo run -p ibrief-app -- sources evolve # score sources + drift watchdog cargo run -p ibrief-app -- research "What's new in local LLMs?"
Tech stack
Rust 1.95 · Edition 2024Tokio asyncOllama local LLMssqlx + SQLitefeed-rs ingestreqwest Telegram APIrand_distr Thompson12 crates · CI green
FAQ
Frequently asked.
What is ibrief?
A self-improving, personalized morning briefing. It builds a curated briefing overnight and adapts daily to measured feedback — fully local in Rust, using local LLMs via Ollama.
How does ibrief improve itself?
Across three safeguarded axes: weights (Thompson bandit), prompts (shadow test) and sources (scoring plus a drift watchdog). Every change passes a safety gate and is versioned and reversible.
How does ibrief avoid a filter bubble?
Engagement is not the optimization target. Two sections — Counterpoint and Wildcard — cannot be disabled, and a drift watchdog enforces source diversity even at the cost of the short-term score.
Do I need a cloud API?
No. ibrief runs fully local with Ollama. A cloud frontier model is optional and only used for periodic judge calibration through an existing subscription (claude -p).