Obi

Obi AI features for existing products

I help support teams test and ship AI. Fixed prices, measured results, everything in writing.

Fixed-price offers

The smallest way to start

$99 bot spot check

I ask the AI assistant on your site or help center 25 questions your public docs can answer. You get a scored sheet: what it got right, what it got wrong, and the five fixes I would make first.

  • Results within 48 hours
  • Public bot only: no data access, no NDA, no call
  • Pay after you see the results, and only if they were useful
Request a spot check

Invoiced through Mellow, for registered companies.

  • AI bot audit

    Already have a support bot? I test it on 100 of your real questions and send a scored report, the main failures, and a reusable test set.

    $1,400 · 5 business days

  • Ticket triage pilot

    Routes tickets or email by topic and urgency, and sends unclear cases to a person. Tested on your labeled examples.

    $2,900 · 7 days

  • Support answers pilot

    Answers questions from your help docs with citations, and hands off what it cannot answer. Tested on real customer questions.

    $3,400 · 10 days

After any offer, I can quote the full implementation at a fixed price. You pay model API costs directly.

Case studies

Tests on public data, with methods and results documented.

  • Benchmark · BANKING77 · Nov 2026

    Decision models vs. LLMs

    Eight models, 1,001 bank support messages, 77 intents. Given only the intent names, four of five decision models beat all three LLMs I tested. Three of the four also cost less.

    Best accuracy,
    Best LLM,
    Per 1,000 messages,
    Routed automatically at accuracy
    Read the test

    Accuracy vs. cost

The 7-day pilot

This is the ticket triage pilot: one workflow, $2,900 fixed, clear numbers. The support answers pilot follows the same steps over 10 days. Everything runs in writing. Plan on 2–3 hours from your team. The schedule assumes your examples arrive by day 2; if they come later, the dates move with them.

  1. Day 1

    Scope

    Written brief. Pick one workflow. Agree what done looks like.

  2. Days 2–3

    Label

    Collect 100–300 real examples, anonymized if needed. Label them together so we agree on ground truth.

  3. Days 4–5

    Build and test

    Build the workflow on part of the data. Score it on examples held back from the build.

  4. Day 6

    Report

    Accuracy, estimated hours saved per month, cost per 1,000 items, failure modes, what needs human review.

  5. Day 7

    Handover

    Code, setup guide, captioned screen recording. The code is yours. Then you decide: stop, or implement.

Questions

Where does my data go?

Your choice. Cloud API for speed, or your own servers if data must stay inside. I document exactly which models and regions are used.

What if accuracy is not good enough?

The pilot report shows exactly where it fails. Often we route low-confidence cases to humans. If it is not worth implementing, you stop after the pilot. With 100–300 examples, the numbers support a go or no-go decision. They are not a production guarantee.

Which models do you use?

I test multiple options and recommend based on your accuracy, speed, and cost needs. I do not lock you into one provider.

Will you sign an NDA?

Yes, before I see any of your data.

How does payment work?

Each offer has a fixed price. The $99 spot check is paid after you see the results. The larger offers are 50% to start and 50% when you receive the deliverables. You pay model API costs directly. Invoices via Mellow, which works for registered companies in 100+ countries.

About

I build AI features into existing products. Remote, solo, and I work in writing. I use AI coding agents, but I own testing and the result. My name is Ivan Orlov; online I go by Obi.

Other work

  • Support assistant. Answers from company documents with citations and passes unanswered questions to a person. Code on GitHub.
  • Service-business site. Next.js. Visitors build orders, see instant price, send requests to Telegram.
  • Translation extension. MVP. Translates pages and chats via OpenAI-compatible API, with batching and caching.
  • Moscow travel-time map. Interactive. Shows reachable areas by time, from OpenStreetMap data. Approximate.

Code for the other projects is private; demos on request.

Work together

Email me one workflow that slows your support team down. I'll reply in writing within a day.

Email me