We turn GenAI concepts into products that survive real users.

Problem

Answering messages while driving is dangerous, and the built-in assistants make it worse: they read entire threads out loud, cannot handle a voice note, and stall on “sorry, I didn’t catch that”.

Product

Messaging with both hands on the wheel. The agent reads out a short summary of what arrived instead of the full text, transcribes incoming voice messages, and sends replies dictated by voice — triggered from the steering-wheel button, with no wake word.

Engineering

The core trade-off is recognition: an on-device recogniser handles short commands offline, while a heavier model handles dictation where accuracy beats latency. Contact aliases are inflected grammatically, muted chats stay silent, and the whole thing runs as a foreground service that survives calls, connectivity loss and reboots.

Kotlin Jetpack Compose Hilt · Room ONNX Runtime VoiceInteractionService Whisper

Problem

The market runs 24/7 and the person holding the position does not. Full automation is the obvious answer and the one most people refuse, because handing an algorithm unsupervised access to their money is a step too far.

Product

The assistant watches instead of acting. It monitors the market against user-defined conditions and, when one fires, proposes the trade for one-tap confirmation. Profit targets are set per position; a paid tier adds automatic take-profit — off by default, enabled deliberately.

Engineering

Real money makes correctness the entire job: partial order fills, funds held inside open orders, reconciliation when the user trades manually on the exchange, duplicate-order protection, and retry paths that never leave a position in an unknown state. API keys are stored encrypted.

Python aiogram exchange REST API SQLAlchemy async PostgreSQL APScheduler

Problem

Consumer AI products rarely die of a missing feature. They die because answers drift after a model update, because a handful of users discover they can make the system say anything, or because the monthly inference bill quietly outgrows the revenue.

Product

Two bilingual assistants with paid tiers: one analyses a conversation and reports tone, balance of initiative and recognised manipulation patterns; the other runs a profiling questionnaire, scheduled content and a standard screening questionnaire, with report export and history.

Engineering

An eval suite sits next to the test suite, so prompt and model changes are checked against expected behaviour instead of judged by reading a few answers. Every message passes a two-tier screen — a cheap model call with a keyword fallback — catching injection in both languages. Per-user daily quotas and global rate limiting keep inference spend bounded.

Python PostgreSQL multi-provider LLM offline STT eval suite rate limiting

Problem

On busy rail routes tickets sell out within minutes of opening, and the seats that actually matter go first. What does become available are returns, appearing unpredictably at any hour.

Product

The service watches for exactly the ticket the user wants — date, route, carriage class, upper or lower berth, specific train — and alerts the moment one appears. The purchase stays with the user by design: it finds tickets, it does not buy them.

Engineering

Browser-driven acquisition against a site with no usable public API, with request throttling and a polling scheduler tuned to stay well-behaved. Matching rules live in their own module and are tested against recorded real-world responses, so a change in the source page fails loudly in tests rather than silently in production.

Python Playwright aiogram async SQLite Docker Compose

Problem

Not every problem deserves a product. Some are small, repetitive and personal — pull the text out of a screenshot, understand a letter in a language you are still learning, justify a conference trip to a manager. Building an app for each is absurd; doing them by hand every week is worse.

Product

Four focused Telegram bots, one job each. A vision bot classifies a photo — screenshot, business letter, chart, slide — and extracts its content accordingly, reading axes and legends where they exist. A translator between Russian and six languages explains the difficult words instead of only replacing them. A European Portuguese tutor returns full conjugation tables and corrects prepositions. The fourth turns a conference URL into a justification memo for a manager.

Engineering

The interesting part is what they share. Each is deliberately small: one model call per task with a strictly expected output, no state beyond user preferences, no framework between the bot and the model. That makes them cheap to run and, more importantly, cheap to abandon — when one stops earning its place, deleting it costs nothing.

aiogram vision models OpenRouter Docker