A plain-language account of Bonaventure: the idea behind it, the four arms, every product with what it does today and what it will do, how a one-person company ships with AI agents, the roadmap, the money, the rules we will not break and the questions we have not answered yet.
Contents
Most company pages tell you what the company wants you to think. This one tells you what the company is. Bonaventure is early: it is being incorporated, it has no revenue yet, and it is built by one founder working with a crew of AI agents. We would rather you learn that from us, in detail, than guess it from a polished landing page.
The document is for three kinds of reader. A potential customer or pilot partner who wants to know what is real before a demo. A potential collaborator or investor who wants to see the whole system and the logic that holds it together. And anyone curious about what an AI-native company looks like from the inside in 2026, including the parts that are still rough.
Everything here comes from the internal documents we actually run the company on: the structure document, the decision log, the monthly plan and the task files. When those change, this document is regenerated. The version and date are on the cover.
Bonaventure is the umbrella for companies that grow without needing human operations.
Mission. Give businesses the system that runs their own growth, and give people transparent access to trustworthy information.
What that means in practice. We build a small number of products for businesses and for consumers. None of them has its own back office. Scanning sources, writing copy, checking facts, sending messages, measuring results, watching budgets: all of that is done by a shared set of engines and runtime agents we call Bonaventure Core. A new product does not get a new server and a new team. It gets a folder in Core.
Why now. In 2026 a single person with the right tooling can design, build, test and ship software that used to need a team of ten. We are not claiming that the software is better than what a team would make. We are claiming that it is good enough to sell, and that the cost structure is so different that products which were never worth building for a small market suddenly are. The Turkish marketplace seller, the regional B2B sales team, the football fan who wants an honest valuation: these are markets a venture-funded team would not touch. We can.
What we are not. We are not a consultancy that happens to have a product, and we are not an app studio chasing a hit. The products exist because they feed each other. That is the whole design, and the next sections explain it.
Bonaventure is a ship's name ("good fortune, good voyage"), and the products carry names from the same maritime family. The structure is one ship with four arms and a keel.
| Arm | The question it answers | What it contains |
|---|---|---|
| Products | What do we sell the customer? | Software and customer-acquisition tools that let a business manage its own growth; consumer-facing apps. |
| Services | Who does the work when software is not enough? | Bespoke AI solutions and operations powered by Core, plus growth consulting. Quote-based, and the first source of cash. |
| Entertainment | What content do the products get? | Games, scenes, music and football data. Each studio feeds Undertow and, through Core, the other products. |
| Core | What runs all of it? | Eight shared capabilities. Product data stays separate; company identity and the engines are shared. |
Products are ports. The lines between them are currents, and there are three kinds: customers and cash, leads and signals, content and data. This is the map that decides whether an idea is allowed on board.
| From | To | What flows | Kind |
|---|---|---|---|
| Ops / Live | Siftah, One | The first customers, who arrive through consulting and move onto a product | customers & cash |
| Ops and first subscriptions | Core | The first cash pays for the engine | customers & cash |
| Çırak | Siftah | A coached seller becomes a Siftah customer | customers |
| Reach Lead | One, Keelson | Companies showing growth signals become leads inside the growth panel | leads & signals |
| Reach Call + Qlead | Ops | A qualified conversation becomes a consulting engagement | leads & signals |
| Undertow | One, Keelson | Aggregate reaction data by region × generation, with district breakdown; never personal (minimum 50 people per cell) | content & data |
| Spindrift, Bergovich | Undertow | Games, scenes and music as cards in the daily feed | content |
| Scoutline | Undertow | Weekly football value cards, through a contract defined in Core, not a product-specific copy | content |
| Core | every arm | Engines, agents, identity and data | core |
For each product: what it is, what it does today, what is planned, what content or data it produces, and the pricing hypothesis where there is one. Status labels mean exactly what they say.
LIVE running with real users or real data BETA running, closed group IN CODE built, not yet in users' hands PLANNED decided and scheduled IDEA not decided
B2C · news and trend cards · the strategic flagship
Undertow is a mobile app that turns the day's news and trends into short, full-screen, playable cards. It is not a social network: there is no user-generated content, no profiles, no comments. You pick a district, and the feed is built for that place.
B2B · customer acquisition · the first product with a price list
Reach finds companies that are showing a growth signal right now, from public sources only, and scores them so a small sales team calls the right people at the right time.
B2B · marketplace seller management · separate product, separate repository
Siftah ("first sale of the day" in Turkish shop talk) is for the small seller on Trendyol who does not want to lose the buy box, does not know which campaign to join and cannot watch prices all day.
B2B · growth platform · One for SMBs and digital brands, Keelson for corporates
One is a growth panel for a brand with its own channels: lists and conversions arrive automatically, journeys actually send at every step, and results are measured against a control group. Keelson is the corporate edition: requesting and executing units, an approval chain, an AI growth team per unit.
B2C · a coach for people who want to start selling on a marketplace
"Çırak" means apprentice. The idea is a guided path from "I have a product" to "I have a store with its first sale", feeding Siftah with its graduates. Not decided, not scheduled.
Entertainment is a studio arm. It does not sell directly; it makes the content that keeps Undertow alive every day and, through Core, gives the B2B products something to say.
Football market values, computed from scratch and explained
Football has one dominant source of player market values, and those values are set by editors and community debate, updated rarely, with no explanation and no uncertainty. Scoutline recomputes the value of every player in seven competitions every week with an explainable model, shows an 80 % range instead of a single number, and tells you the three to five drivers behind each value.
Game studio · AI-native, small, daily
Spindrift (the spray the wind tears from a wave crest, the surface twin of the undertow) is the game studio. We do not compete with hyper-casual studios and their user-acquisition budgets. Our games are 30-to-90-second sessions built from five templates (tap-and-time, swipe-and-choose, sort, guess, rhythm), produced largely by agents from the day's news and trends, and shipped as single HTML files that need no store update.
Music studio
Bergovich makes the music. A 20-plus track album produced with a generative music service under a commercial licence is being prepared for streaming release; the tracks play inside Undertow's rhythm games and Explore tab and in Spindrift's games, and the games bring listeners to the tracks. The copyright status of AI-generated music is unsettled, and that is on our lawyer's list. It does not stop the work.
Services are where the first money comes from, and where the first product customers come from. There is no separate services website; the main site carries the offer.
Consulting packages and bespoke AI solutions built on Core: process automation for an SMB, an agent set-up, a content or reporting engine, growth consulting, and operations takeover (we run a function for you). Each engagement is scoped against Core first, so the customer gets an engine that already exists and is already tested, not a one-off script. An Ops customer is the natural first user of Siftah or One.
Live consulting for your operations and your AI needs, billed by the minute with a credit model. Hourly rates and service areas are being set; the market range for SMB digital-marketing consulting in Turkey is roughly 1,500–4,000 TL per hour and we will publish our rate rather than quote case by case.
Core is the keel. Nobody buys it, everything stands on it. It is not one program; it is eight capabilities that today live in different repositories and are being pulled together one at a time. A product gets a folder, a budget counter and a set of contracts. Product data stays in its own store; what is shared is identity, the engines and the agents.
| Compartment | What it is | Where it runs today |
|---|---|---|
| Automation engine | A workflow engine on its own server; one flow folder per product; shared error, log and budget counters. Every flow writes to a queue, so if the engine falls, products keep running from their last state. | n8n in the One/Keelson platform; 53 flows |
| Agent crew | Six lines, nineteen runtime roles: crawler, deduplicator, source warden; writer, variant maker, question master; verifier, compliance, sensitivity; voice agent, qualifier, call analyst; reporter, measurer, insight patterner; and the wardens: budget, anomaly, complaint, ad ceiling. Model rule: the cheapest model that is good enough, rules without a model where possible. | Crawling, writing, auditing and the wardens are live in the news engine; voice is closed pending legal review; analysis exists in code |
| Crawling engine | Scheduled source reading with robots.txt compliance, one request per source per run, source health scoring, dead-and-stale detection, clustering of the same event. | News engine (around 50 sources, every 30 minutes) and Reach Lead (242 recipes); Scoutline reuses it for club sites |
| Voice engine | Ninety-second AI calls with an approved script, a qualification rule and a summary from the recording. | Closed until the legal opinion on AI calling arrives |
| Delivery | Sending: e-mail, SMS, push, with consent registry checks and per-channel ceilings. | In code in One/Keelson; needs provider accounts that come with incorporation |
| Identity | One sign-in contract for every panel: password hashing (scrypt, with a PBKDF2 fallback), signed session tokens, a common cookie, rate limiting, a shared role dictionary, and conformance vectors that both the Node and the Python implementation must pass. Single sign-on across products comes after the first paying customer. | Node package live in One/Keelson; Python package in progress for Reach and Siftah; Undertow stays on its own auth provider with a bridge later |
| Data and Insights | Aggregate, never personal: region × generation × district × card type, minimum 50 people per cell, rows below that fold up to city level. Explainable scoring (the Scoutline factor chain is the pattern) for every product that scores anything. | Partly in code; the Scoutline score pattern is the reference implementation |
| Bridge | Internal control: the dashboard that shows what every product and every agent is doing, what it cost, what is open, and what the human needs to decide. | Live for the news engine (daily cards, topics, diversity, source health); company-wide board live internally |
Bonaventure has one founder. The founder does not write code. The work is done by a crew of AI coding agents, each with a role, and coordinated by a senior model that acts as architect, reviewer and keeper of the decision log.
| Who | Role | Rule |
|---|---|---|
| Coordinator (Claude) | Writes the briefs and task files, reviews every pull request, maintains the structure document, the decision log and the plan, keeps the architecture consistent across repositories and stops duplicate work. | Writes no product code beyond one-line fixes. Everything goes through GitHub. |
| Builder agents (three different vendors) | Application screens, Python engines, websites, backend and database, deployment, games and visuals. Each agent has its own clone of each repository. | One open task per agent, one open pull request per product. A fabricated test result means the pull request is rejected. |
| Founder | Merges pull requests, makes decisions, holds every account and every secret, talks to customers. | About thirty minutes a day plus two decision blocks a week. Every decision gets a number in the log. |
The loop. A need is placed in the structure (which arm, which current). The coordinator writes a task file with a file boundary, acceptance criteria and an effort estimate in senior-developer hours. An agent takes a branch, does the work, opens a pull request with the four mandatory sections and the number of tokens it used. The coordinator verifies locally (runs the tests, checks the diff against the boundary, scans for real names and secrets) and comments. The founder merges. The structure document and the plan are updated in the same commit as the decision that changed them; otherwise the decision is not considered closed.
What this looks like in numbers. Across the ten repositories, since the first commit in late October 2026: about 150 merged pull requests, roughly 1,100 senior-developer-hours of equivalent work, and 94 test suites with more than 3,500 checks on the largest platform alone. On a good night the crew ships ten pull requests while the founder sleeps, and reviews them before breakfast.
We track three numbers for every task: the effort it would take a senior developer (set by the coordinator from the size and nature of the change), the tokens the agent actually used (reported in the pull request, estimated by a rule if not), and the cost of those tokens at the vendor's list price. From that we get a cost per hour of equivalent work for each agent, and what the whole body of work would have cost if it had all been done by the most expensive agent.
| Measure (31 October 2026) | Value | Note |
|---|---|---|
| Equivalent effort shipped | about 1,100 senior-developer hours | Sum of effort estimates on merged work; the estimate rule is being calibrated as real token counts accumulate |
| Token cost at list price | about 10,000 $ | Most of it is the coordinator's review and design work; the actual bill is lower because of subscription plans and cached context |
| Cheapest builder | about 0.01 $ per equivalent hour | A low-cost model that does most of the volume work under tight briefs |
| Most expensive | about 41 $ per equivalent hour | The coordinator; worth it for review and architecture, wasteful for volume |
| Saved versus doing everything at the top rate | about 36,000 $ | And versus a human team the gap is an order of magnitude larger; we do not publish that number because we cannot measure it honestly |
The point is not that the work is cheap. The point is that the cost curve is flat enough that products for small markets, run by agents with a human on approval, can be profitable at prices those markets can pay. That is the bet.
Running costs today are deliberately small: the news engine on one small cloud machine (a few dollars a month plus a capped AI budget of about one dollar a day), static sites on a free tier, a sports data licence at 19 $ a month, domain names, and the AI subscriptions. Every proposal that adds a monthly cost has to state it.
| # | Line | Time to first payment | Depends on | Unit revenue |
|---|---|---|---|---|
| 1 | Bonaventure Ops / Live | 1–2 weeks after the services section and the rate are published | Founder's hours; an invoicing entity | Hourly rate, to be published |
| 2 | Reach Lead | 4–6 weeks: demos in November, decision gate at month end (5 demos, 3 "I'd pay"), first invoice in December | Ten decision-makers in the target vertical; the euro-invoicing question | From 299 € / month |
| 3 | Siftah | 8–12 weeks: pilot seller in December, first subscription in January | A pilot seller with a real Trendyol account | Subscription, hypothesis 4,900–14,900 TL / month |
| — | Undertow | Not in the revenue order | Strategic visibility, not cash. It stays at the top of the company's priorities for that reason. | |
| — | One / Keelson | After incorporation | Corporate sales are slow; provider accounts need a legal entity. | |
Undertow Plus in month three after launch. Core identity as a single sign-on service after the first paying customer. Reach multi-tenant hosting at ten customers. Siftah consignment after three satisfied sellers. Android for Undertow. Reach Call and Qlead if and when the legal opinion allows. Çırak if Siftah proves the path.
We keep a decision log with 41 numbered decisions so far, and a list of what is still open. The open items that matter most, stated plainly:
| Question | Why it matters | Status |
|---|---|---|
| Incorporation and partnership | The company is not yet a legal entity; a verbal partnership understanding exists and is flagged internally as a risk to be put in writing. | With the lawyer; decision on timing and place pending |
| Founder's existing employment | Side-activity and intellectual-property terms have to be confirmed in writing before certain lines (and the founder's name) appear publicly. | With the lawyer; this is why no founder name appears here |
| News publishing rights | Many Turkish news sources forbid reproduction. Our model is headline + own summary + link, but public launch of the news feed waits for an opinion. | Gate before public launch; sources approved one by one |
| AI voice calling | Reach Call and Qlead are closed until the legal position on automated calls is clear. | Waitlist only |
| Scoutline and personality rights | Publishing a value under a real player's name, correction and removal, the data licence's derivative-work and mass-media clauses, distance from betting. | Eight questions filed; synthetic data until answered |
| Euro pricing for a Turkish company | Showing euro prices and invoicing in lira may conflict with currency rules. | With the lawyer; euro prices stand meanwhile |
| AI-generated music | Copyright protection for generated works is unsettled. | Asked; not blocking |
| Does anyone pay? | Reach's November gate (5 demos, 3 "I'd pay") is the first real test of the whole thesis. | Open until 30 November 2026 |
Risks we watch and the early signal for each: a data licence that forbids derived valuations (we read the terms before we model); a model that is worse than the incumbent (three months of silent validation before any claim); no traffic for Scoutline (it does not grow without the weekly list and the Undertow cards); scope creep ("just one more league": seven competitions are locked until validation is published); and the founder's time, which is the real constraint on everything above.
| Name | Meaning |
|---|---|
| Bonaventure | A ship's name: good fortune, good voyage. The umbrella. |
| Keelson | The inner beam that strengthens the keel. The corporate edition that sits on the platform. |
| Reach | A point of sail; also the act of reaching a customer. |
| Siftah | Turkish shop talk for the first sale of the day. |
| Undertow | The current beneath the surface. The news under the news. |
| Spindrift | Spray torn from a wave crest by the wind. The surface twin of the undertow. |
| Scoutline | A scout's line: the value curve of a player, drawn by a model instead of a rumour. |
| Bergovich | The music studio's name. Just a name. |
| Çırak | Apprentice. |
| Kayık, Sandal | Rowing boat, dinghy: the Reach package names, smallest first. |
| Bridge | Where the ship is steered from: the internal control panel. |
Contact: [email protected] · bonaventurecorp.com
Company record: Bonaventure is being incorporated; the legal entity and company details will be added at incorporation. Until then nothing in this document is an offer, and every price is either a published introductory price (Reach) or an explicitly labelled hypothesis.
This document is generated from the internal structure document, decision log and plan, and regenerated when they change. Version 1 was produced on 31 October 2026. If you find a claim here that does not match what you see in a product, tell us; that is a bug, and we fix bugs.