Three weeks of shipping Vyneron solo: redesigned app, central Telegram bot, natural-language dates in 6 languages
I'm three weeks out from a Product Hunt launch and wanted to share
what a focused solo sprint actually looks like — the good, the
weird, and the bugs I had to chase down myself.
Headline numbers: ~30+ deploys, 1,037 → 1,225 passing tests, every
core view in the app rewritten. Solo, no team, no contractors.
A few things I'm proud of:
→ Full app redesign GA'd. Home, tasks, routines, notes, archive —
all new layouts on top of a shared design system. Last week I
retired the legacy screens and feature flags entirely: 338 lines
of dead code removed, 1,225 tests still passing.
→ Central Telegram bot. Removed the per-user bot-token setup that
was costing me ~20% of new signups. Telegram is now an optional
super-power, not a setup hurdle. Web-first stays primary.
→ Deterministic date parsers. The AI was doing weekday math
stochastically — same input, different off-by-N every call. Wrote
a regex-based parser that handles "10 minutes later", "next Monday
at 10", and TR locative suffix forms ("Pazartesi 10 da") in 6
languages. Pre/post-AI override pattern: let AI handle intent, let
deterministic code handle dates. Won't repeat that mistake.
→ AI photo → tags. Send a photo of a handwritten note, whiteboard,
or receipt — Gemini Vision reads it and suggests tags inline. The
note stays searchable months later without anyone manually tagging
anything.
→ Bring-Your-Own-Key, now with 4 providers. Gemini, OpenAI, Groq,
and Anthropic Claude. 0% markup on top of provider pricing. If
you're already paying OpenAI or Anthropic, plug the key in and the
AI side is effectively free.
The honest part: I shipped a migration drift bug to prod last week
because I bypassed deploy.sh and forgot a new schema file in my
tarball. Scheduler crashed for 8 minutes. Fixed, committed the
lesson to my own playbook. Solo means you're the only person who
catches your blind spots.
Next 3 months: AEO authority work, comparison content, and the PH
launch itself. The product is solid; what's left is making sure the
right people hear about it.
If anyone here has shipped a PH launch as a solo dev — what
surprised you most in the final stretch?
This feels like the right time to tighten the positioning before Product Hunt.
The product is no longer just a task app. From what you described, it is becoming a personal workflow layer: tasks, routines, notes, Telegram capture, deterministic date handling, AI photo tagging, and bring-your-own-key AI providers. That is much stronger than “AI productivity app,” but the launch copy needs to make that category clear fast.
The biggest PH risk is that people see a long feature list instead of one sharp promise. I’d frame it around reliable capture and retrieval for people who run their life through scattered inputs: messages, notes, screenshots, receipts, routines, and dates.
Naming is also worth pressure-testing now. Vyneron is distinctive, but it does not immediately tell me what shelf to put the product on. For a workflow/productivity platform with AI underneath, a cleaner name like Xevoa.com would probably carry the broader direction better and feel less like something that needs explanation before the product does.
Really like the “AI for intent, deterministic code for dates” pattern. In conversational workflows, I’ve seen the nastiest bugs show up exactly where the system needs to be predictable: time, language, names, follow-ups. Did you move date parsing fully out of the AI layer, or do you still keep AI as a fallback?
Hybrid. We pulled the arithmetic fully out of the AI layer — three deterministic parsers run before the LLM sees the message: relative time ("in 3 min", "2 saat sonra"), weekday+time ("Monday 10am"), and title cleanup (the matched time span gets cut out of the title so it can't drift back as "Meeting at 3pm").
The LLM still owns intent (task vs note vs routine, priority, content) and anything the parsers don't match ("end of the month"). If a parser fires, its due_date overrides whatever the AI returned. AI for semantics, deterministic for math — letting the LLM do arithmetic was our single biggest source of "why is this scheduled next year" bugs.