00 / IndexClose

Nuits, a one-person AI design and engineering studio in Helsinki.

Grainy abstract composition of arches and horizontal bands with a small figure standing in a lit doorway

I design and build AI products for companies.

From the first sketch to working code, by one person.

I work out what the AI should do, design the screens and write the code. Everything ships with fallbacks, because models get things wrong.

  • 5x design award winner
  • 20+ clients
  • Projects of two to six weeks

Email meBook a call

Replies within 48 hours.

02 / Projects

Things I designed and built.

A product I built for a client, products of my own, and three companies I have worked with.

Rows of banded spruce trees lining a straight path, a small figure walking in
  1. Fluence Sync, its live page
    • Daily data sync
    • Role-based permissions
    • Session recordings and transcripts
    • Competency assessments

    Livev1, then v2Built for Fluence

    Fluence Sync

    A dashboard for clinical training teams. It shows trainee progress, assessments and recorded sessions in one place, instead of spreadsheets and scattered reports.

    A custom AI tool reads each roleplay session and returns a summary, scores and recommendations in real time. Trainers review the recording with the transcript next to it.

    I designed and built version one myself. Version two came with a small team, once it had earned one.

    Open Fluence Sync
  2. Mochi, its live page

    Waitlist2026

    Mochi

    A Mac app that turns what you say into notes in your Obsidian vault.

    Speech recognition and speaker separation run on your Mac, so no audio is uploaded. The exact transcript is saved first, then a language model tidies a copy. If the model is down, you still have the note.

    • 0 bytes of audio uploaded
    • 4 speakers separated
    Open Mochi
  3. SignalScout, its live page

    Live2026

    SignalScout

    Turns a CSV file of sales leads into opening lines for cold emails, each with a source, a date and a confidence level.

    AI research agents look for evidence about each lead, and a model writes the opening line from it. If the research finds nothing, the line stays blank instead of made up.

    Open SignalScout
  4. Lexborn, its live page

    Pre-launch2026

    Lexborn

    Speaking practice for English speakers learning another language. Seven languages, three conversation partners each.

    Realtime voice models hold the conversation and remember you between sessions. Corrections wait until the end, grouped next to the sentence you meant. Words you could not find become flashcards.

    Open Lexborn
  5. Cringe, its live page

    Waitlist2026

    Cringe

    A dating app designed to get you off the app and outside.

    A model rewrites your profile to be honest rather than flattering. It can be blunt about your claims, never about you. Some profiles are AI plants, and the app labels them.

    Open Cringe
  6. Doveborn, its live page

    Waitlist2026

    Doveborn

    Two people write each other one letter a week. Neither letter opens until both are sealed.

    Optional writing prompts come from a model that never sees what either of you wrote. That privacy rule came first, and the feature was built around it.

    Open Doveborn
  7. BapsOS, its live page

    Live2026

    BapsOS

    Windows 98, rebuilt to run in a browser tab. A desktop called Fen98 and 27 working applications.

    One of the 27 apps is an AI chat. The others include a spreadsheet with formulas, Paint and Minesweeper. The hardware specs are made up, and very confident.

    • 27 working applications
    Open BapsOS
  8. Echoborn, its live page

    Live2026

    Echoborn

    A musical instrument in your browser. You play it by pinching the air in front of your camera.

    Hand tracking runs on your device and nothing leaves it. No language model, no server, no account. Low latency is the whole product.

    Open Echoborn
  9. UX Design GPT, its live page

    Live2023

    UX Design GPT

    A custom GPT that answers UX design questions from six years of my design practice. Pay what you want.

    Built on my own written design knowledge. It asks about your constraints before it answers, and says when a method is overkill.

    • 758 sales
    • 4.8 from 17 reviews
    Open UX Design GPT
  10. PokeGPT, its live page

    Live2023

    PokeGPT

    A custom GPT connected to a real API. I built it days after GPT Actions launched, to test whether they worked.

    An OpenAPI schema and a set of instructions. The model has to fetch each record instead of guessing it, so wrong answers showed up in seconds.

    • 21 sales
    Open PokeGPT
  11. Window, its live page

    Live2026

    Window

    An online store with 350 products and nothing for sale. I built it for my mother, who could not stop shopping online.

    No AI. I photographed all 350 products myself, mostly at night. Generated images would have been faster, and the store would have looked like a joke.

    • 350 products
    • 0 things for sale
    Open Window
  12. Fablebloom, its live page

    Pre-launch2026

    Fablebloom

    A relaxed strategy card game in the browser. No hidden cards, no timers, no card packs.

    No AI in the game. I balanced all 84 cards by hand, which took longer than asking a model.

    Open Fablebloom
  13. Music by Baps, its live page

    Live2026

    Music by Baps

    The website for my music label and its three releases. It has more WebGL than three songs need.

    No AI. I made the songs, the covers and the shaders. The songs are about sweeteners and compliance.

    Open Music by Baps
  14. Primobots, its live page

    Shipped2021

    Primobots

    A Web3 AR and VR startup I co-founded. 5,555 3D collectibles, a Snap Lens and full-body avatars, built over fourteen months.

    No AI. In 2021 the models could not draw yet, so I built the pipeline in Blender and shipped it by hand.

    • 5,555 3D collectibles
    • 14 months
    Open Primobots
  15. Fluence, its live page

    Ongoing2024 to nowSenior AI and design lead / Healthcare education, New York

    Fluence

    I take new AI products and workflows from idea to launch, Fluence Sync among them. I also built the brand and the website, on Framer and Shopify.

    Open Fluence
  16. Blockdaemon, its live page

    Past2022Design and marketing / Fintech infrastructure, Los Angeles

    Blockdaemon

    I designed for new and existing protocols, and made brand identity work used at 25+ events a year.

    Open Blockdaemon
  17. OpenClassrooms, its live page

    Past2020 to 2023Mentor and career coach / EdTech, Paris

    OpenClassrooms

    I mentored designers and developers in Europe, the USA and Japan until they finished their projects.

    Open OpenClassrooms

03 / How it works

Four steps, in order.

01DiscoverI learn what success looks like from the people who will judge it.
02Design and testI prototype how the AI behaves, and your team tries it early.
03Build and connectI build the product, connect it to your systems and add monitoring.
04Launch and improveI launch it, tune the prompts and leave you a written playbook.
Terraced bands with a long staircase rising to a lit doorway, a small figure climbing

04 / Ways to work

Four kinds of project, two to six weeks.

ProjectWeeksWhat you get
AI product design3 to 5The user flows, the prompts, what happens when the AI fails, and how to test it.
Automation and workflows2 to 4A map of your workflows, a plan to automate them safely, runbooks and monitoring.
Design systems4 to 6A component library with states and tokens, patterns for AI content, accessibility checks.
Product foundations3 to 6The groundwork for AI features: architecture, data and API contracts, quality checks.
A grid of square planters with round pools, seen from above, a small figure among them

05 / A client's words, for once

A building facade as a grid of dark windows, one lit with a figure at a desk

Bapusaheb delivers with rare clarity: crisp flows, production-ready specs, and the rigor to make AI features safe and reliable.

Product leader, US client who asked not to be named.

06 / Prompt or punt

Prompt or punt.

The AI does it, or a person does. Ten calls.

A winding snowy path between spruce, with a signpost whose arms are blank and a small figure on the path

Example

Summarise every support ticket from the weekend so Monday's standup takes ten minutes.

The call: prompt.

  1. Summarise every support ticket from the weekend so Monday's standup takes ten minutes. The call: prompt. A summary someone reads before acting on it is the job.
  2. Draft the release notes from the merged pull requests. The call: prompt. A draft an engineer reads before it ships.
  3. Write the dosage instructions on the medication screen and ship them with the next release. The call: punt. A wrong number hurts someone, so a pharmacist writes it.
  4. Pick which three people to let go from the support team. The call: punt. Someone has to own that and say it to their faces.
  5. Price the three-year enterprise contract for the bank. The call: punt. A price is a promise with a signature on it.
  6. Tag the 40,000 product photos in the archive by colour and category. The call: prompt. Bulk, dull and checkable against a hand-tagged sample.
  7. Write the public apology for yesterday's four-hour outage. The call: punt. The model can draft the timeline, but it cannot be sorry.
  8. Choose between the two candidates for cofounder. The call: punt. A judgement someone lives with for ten years.
  9. Run a first-pass translation of the help centre into Finnish and German. The call: prompt. First pass, and a native speaker reviews it before it goes live.
  10. Freeze any customer account the fraud score flags, with no review. The call: punt. A frozen account is someone's rent, so a person decides.
  11. Generate alt text for the 3,000 images in the blog archive, with an editor checking a sample. The call: prompt. Better than the empty attribute, and the sample catches the worst.
  12. Write the new terms of service for the payments launch. The call: punt. A contract with every customer at once, so a lawyer owns it.
  13. Write the toast message that confirms a file was saved. The call: prompt. Low stakes, and there are only so many ways to say saved.
  14. Pick the senior designer to hire from the final three. The call: punt. A hire someone has to explain a year later.
  15. Triage the inbound email into sales, support and spam. The call: prompt. Classification, with a queue for anything unsure.
  16. Approve refunds under 20 euros on orders that match the written policy. The call: prompt. Small, capped and written down, so give it rails and a log.
  17. Decide the 6,000 euro refund for the customer whose lawyer emailed this morning. The call: punt. A lawyer in the thread means the decision needs a name.
  18. Turn the recorded sales calls into notes in the CRM. The call: prompt. Notes a rep reads before the next call.
  19. Tell patients in the app whether their symptoms need a doctor. The call: punt. A body on the other end needs a clinician's name.
  20. Generate 10,000 rows of realistic fake customers for the staging database. The call: prompt. Nobody is harmed by a fake customer called Pirjo.
  21. Leave first-pass review comments on every pull request before an engineer looks at it. The call: prompt. A second reader that never gets bored, and a person still merges.
  22. Set the monthly price of the new Pro plan. The call: punt. A bet someone answers for when the churn numbers arrive.
  23. Find duplicate bug reports in the backlog and suggest merges. The call: prompt. Suggest is the word, and a person clicks merge.
  24. Let the agent delete stale customer accounts on its own. The call: punt. Deleting cannot be undone, so a person presses the button.
  25. Read the 200 supplier contracts and list every auto-renewal clause with its notice period. The call: prompt. Extraction from a pile, checked against a sample.
  26. Write the annual performance review for each person on the team. The call: punt. Pay and careers, so the manager stands behind every line.
  27. Write first drafts of 1,200 product descriptions for the shop. The call: prompt. Bulk drafts against a style guide, approved in batches.
  28. Answer questions in the chat on the pricing page, and hand over to a person after two failed replies. The call: prompt. A scope, rails and a way out.

I design and build AI products for companies.

Send me the one you need built.

Write a few lines about the product and your deadline. I will tell you whether I can help, and how.

hi@brandnuits.com

Replies within 48 hours

Nuits is a one-person AI design and engineering studio in Helsinki. Projects run two to six weeks. Email hi@brandnuits.com, replies within 48 hours.

A frozen lake under concentric rings of sky, a small figure standing on the ice
Pine islands on calm banded water, and a small figure in a boat.

This site was designed and built by me, as usual.

  • Built with Next.js, React and Motion, hosted on Cloudflare.
  • The sky behind this page is a WebGL shader.
  • Set in Familjen Grotesk and IBM Plex Mono.
  • Analytics: self-hosted OpenPanel in the EU, no cookies.