An AI secretary for $77 a month: 88% of answers from DeepSeek, 77% of the budget goes to Opus

Over six months, 118 users sent Liza Petrovna 8,568 queries. The bot lives in Telegram and MAX, understands text, voice, and photos, and either answers on its own or delegates to a doctor, a lawyer, a navigator, a designer. Memory is unified: context isn't lost between agents, between days, or between messengers.

Over the last 30 days, 2,500 replies cost $77 total. Of that, $51 came from the secretary itself, with the rest going to specialists. The key figure: 7% of queries go to Claude Opus and consume 77% of the budget ($39). A single Opus reply costs as much as 50 DeepSeek replies.

Routing instead of intelligence

By default, Liza answers through DeepSeek. Opus only kicks in on a trigger. The first: a regex rule matching complexity keywords — "analyze," "calculate," "plan," "break down," "think it through carefully," "explain in detail." The second: a message with context longer than 2,500 characters, signaling the person is stuck and wants a solid answer.

Web search through an in-house SearXNG instance is available to every model, so it's not a reason to escalate. A regex rule beats a model-based classifier: it's fully visible, testable, and explainable. A classifier would add latency and cost to every single reply just to make decisions about that 7%.

The models in numbers

A 30-day breakdown (2,500 replies):

• DeepSeek (text + vision): 88%, $0.002-0.007 per reply, 2.5-12 sec • Claude Opus 5: 7%, $0.21-0.24 per reply, 18-38 sec • Claude Haiku 4.5: 2.4%, $0.006 per reply, 43 sec • GLM 5.2 (experimental): 2.2%, $0.026 per reply, 5 sec

Total budget: $51 a month. Opus eats $39 of it. The whole scheme boils down to: cheap by default, expensive only rarely.

What people actually ask

All 8,568 queries were classified via DeepSeek in batches. 21% turned out to be filler like "go on" or "you there?" That left 6,796 real tasks by topic:

• Parse a photo, document, or screenshot — 19.6%
• Make an image, video, or design — 12.5%
• Computer, IT, setup — 9.2%
• Find a fact, news item, or reference — 9.1%
• Money, business, work — 7.1%
• Equipment, cars, aviation — 6.9%
• Health — 6.2%
• Legal and government questions — 5.6%
• Chat, life advice — 5.0%
• Write or translate text — 4.9%
• Calendar and reminders — 4.7%
• Trips, places, restaurants — 4.7%
• Studies, languages, sports — 2.5%
• Family, home, household — 2.1%

An AI secretary that's anything but a secretary: calendar and reminders make up just 4.7%. Every fifth task involves a photo — without vision, the bot would lose its core use case. Images and video are requested 2.5 times more often than text.

Mistakes and defenses

Liza confuses photos within a long context, latching onto the previous image. Now, each new message is set off with a === NEW MESSAGE marker. She also gets dates wrong, failing to understand "today" without an explicit substitution, and misdetects language, apparently weighing metadata over actual text.

She's protected against attacks: she rejects prompt injection in Russian and base64, doesn't trust claims of administrator status, won't record fabricated rules, and refuses to invent laws that don't exist.

Memory runs on pgvector: a hybrid search combining semantics and word matching, with a median response time of 17 ms. Embeddings are computed by a local service running deepvk/USER2-small, so replies don't wait on it — if the service goes down, search falls back to word matching. Each person has their own memory: 7,933 episodes from 97 people, all vectorized. Third-party identifiers are stripped from the text. Deleting a user deletes their history.

Liza is free: t.me/LizaPetrovna_bot.