Alexandria: An Autonomous Personalized Book Curator That Learns My Taste - P1 System Design

1. What Alexandria is, and what it isn't

"Autonomous" here means it runs every day without me: it wakes on a schedule, does the same five stages in the same order, and stops. It is not an open-ended agent that decides its own next step. That choice makes it cheap, testable, and easy to replay when a pick goes wrong.

2. The taste registry

Taste is data with a timeline. Every change is an event, so Alexandria can answer "what did I like in March?" and the learning loop can tell a real shift in taste from a bad week.

3. Stage 1: Discover

Alexandria crawls the sites I track: bookstores and publishers I register in a web form, the same way I register tastes. A wide crawl every Sunday builds the candidate pool; a small daily crawl picks up new releases. No paid data APIs.

The form takes a site URL, its kind (bookstore, publisher, prize, review), the pages to watch (new releases, a category), and the language. Before the first crawl, Alexandria reads the site's robots.txt and sets its mode: crawl, or link-only (no fetching; store search links only). The seed list below is preloaded, with modes from robots.txt checked on Nov 2, 2026.

Everything is stored per work (one book, many editions), keyed by ISBN-13 or normalized title and author, so the same book found on five sites is one candidate.

4. Stage 2: Shortlist by taste

From a few hundred candidates to five, in three passes: hard filters, a cheap score, then an LLM judge on the top twenty.

Which taste gets today? A slot planner picks the facet before the book. It uses Thompson sampling: each facet keeps a Beta distribution of "how often I liked its picks", seeded from my priorities, and the planner samples from each and takes the highest. Math and physics history start ahead, but a facet that keeps getting "Not for me" loses slots on its own. One day a week is a labeled wildcard outside my top facets, so the loop doesn't shrink my taste to what it already knows.

5. Stage 3: Review check

The top three of the shortlist get a full review check. The best of the three after reviews becomes today's pick; the other two are kept as runner-ups.

The comparison of the three books is stored too: which one won and why. That gives a ranked group every day, the same group-relative signal GRPO learns from in the Wikipedia agent tutorial.

6. Stage 4: Recommend, one book a day

At 07:00 ICT the pick is published as a page on this site. For the proof of concept that page is the only channel, and it is also where I react. Telegram comes later.

7. Stage 5: Recognize my reaction

Only four signals, all on the daily page. Each becomes a row in reaction, then a number in reward. Rewards can be revised: finishing a book weeks later updates the reward of the day it was recommended.

A comment such as "great story but too equation-heavy for a morning read" is parsed by a small model behind the gateway into structured fields: sentiment, aspects (narrative +, math density −), and a suggested change ("lower math depth for weekday picks"). Suggestions never change the registry by themselves; they wait on the console page for one-click confirmation.

Nothing else is recorded: no purchase tracking, no reading-time guesses, and silence is not treated as a reaction.

8. The learning loop, like RL but sized for one reader

One reader produces about 365 labels a year. Full RL on an LLM would starve on that. So Alexandria climbs in levels, and each level only switches on when there is enough data to beat the one below it on the offline eval.

Weekly reflection. Every Sunday, the reflection model reads the week's reactions and comments and proposes a profile diff in plain words ("you liked narrative physics history more than technical treatments; lower math depth on weekdays"). I approve or edit it on the console page, and it becomes a new profile_version.

9. Tech stack

A scheduled Python pipeline, one Postgres database, a few small edge functions, an AI gateway, and pages on this site. Nothing to keep running except the database.

The AI gateway

Every AI job (extract, judge, digest, write, parse a comment, reflect) goes through one function: gateway.run(task, input, schema). It returns typed output plus the model used, tokens, cost and latency. Which model serves which task lives in one config file, so switching providers is an edit, not a code change.

judge:      { model: gemini-3.8-flash,      fallback: [deepseek-v4-pro, claude-sonnet-5.5] }
digest:     { model: gemini-3.8-flash,      fallback: [deepseek-v4-pro] }
write:      { model: gemini-3.8-flash,      fallback: [claude-haiku-5.5] }
extract:    { model: gemini-3.1-flash-lite, fallback: [openrouter/free] }
parse:      { model: gemini-3.1-flash-lite, fallback: [openrouter/free] }
reflect:    { model: claude-sonnet-5.5 }
budget:     { usd_per_day: 0.50, on_exceed: free_routes_only }
shadow:     { judge: deepseek-v4-pro, share: 0.2 }

Choices and why

Data model

10. Guardrails

11. Cost estimation

Crawling and scheduling run inside free tiers, so the only real cost is the AI model. Depending on the route through the gateway, that is $0 on free tiers or about $3–8 a month on paid models.

Daily AI workload

AI API comparison for this workload

Monthly cost of the heavy jobs (30 days × 64k in / 13.7k out) at each model's standard price. "Tier" is where the model sits in its own provider's line-up; how well a model actually picks books for me is measured by the gateway's shadow runs, not by this table.

Four routes

Plan: start on route A, shadow-run route B or C on 20% of judge calls (well under $1 a month), and let the replay eval decide after a month. DeepSeek's off-peak price applies to the nightly run: 00:00 ICT is 17:00 UTC, outside its peak hours.

Everything else: GitHub Actions (about 10 minutes a day, inside the free monthly minutes), Supabase free tier, GitHub Pages, and embeddings on Gemini's free tier. No paid data APIs.

Prices as of Nov 2, 2026: Google and DeepSeek from their own pricing pages, Anthropic from its price list, and OpenAI, Moonshot (Kimi) and Mistral from OpenRouter's public catalog. Gemini prices above change on Dec 1, 2026 and Jan 1, 2027, as noted. Check before switching routes.

12. Open decisions

Decided items are recorded so the design stays honest about what changed. Open items have a recommended default.

1. Decided Channel. The proof of concept publishes a page on this site and collects my reactions there. Telegram comes later.
2. Decided Editions. Show both the original and the Vietnamese edition when a translation exists.
3. Decided Models. Through the AI gateway, compared by shadow runs and the replay eval.
4. Decided Formats. Any format; no filter on print, ebook or audiobook.
5. Open Starting route. A (free), B, C or D from section 11? Recommended: A, with B shadowed.
6. Open Seed sites. Keep the nine seed sites and their modes, or add others now? Recommended: keep, add more from the form later.
7. Open Where pages live. A separate "Book suggestions" list plus the latest three on the blog index, or every page in the blog index? Recommended: separate list.
8. Open Publishing. Public at 07:00 with a Hold button on the console, or a private preview I approve first? Recommended: public at 07:00 with Hold.
9. Open Monthly cap. Recommended: $10 hard cap, which only matters once a paid route is on.
10. Open Delivery time. Recommended: 07:00 ICT.