Wispr Flow: Dictation That Works EverywhereStop typing. Start speaking. 4x faster.
Sea.rho7
"The AI tutor that works without WiFi"
20 followers
"The AI tutor that works without WiFi"
20 followers
"Sea.rho7 is a hybrid AI platform with two sides. For students (Class 6-12, CBSE/TBSE/ICSE): AI tutoring that keeps working with zero internet. For builders: Sea.rho7 Professional runs Gemini, Codex, and Claude together on every question — one drafts, another adds precision, a third synthesizes the final answer — plus an AI Marketplace directory for every other tool you need. Built solo from Agartala, Tripura, using AI-assisted development."
Interactive
Free Options
Launch Team / Built With
Solo dev building something this ambitious from Agartala is impressive. One thing I want to understand better: when you say the three models collaborate on a single question, how do you handle the latency hit? Running Gemini, Codex, and Claude sequentially on every prompt seems like it would add real wait time, especially if the user is paying API costs on top. Is there caching or parallel routing happening behind the scenes?
@denizbyktedls6 "Great question, and honestly — right now it's sequential, not parallel. Gemini drafts, then Codex does its pass, then Claude synthesizes both. No caching yet either. So yes, there's real latency — typically 7-15 seconds depending on the question, and real cost across all three calls.
The obvious next optimization: Gemini and Codex don't actually depend on each other, so they could run in parallel — only Claude's synthesis step genuinely needs to wait for both. That's next on my list.
Being upfront — this is a solo build, and I've prioritized 'does the collaboration genuinely produce better answers' over latency so far. Feedback like this tells me what to fix next."
solo-built from Agartala, that's genuinely cool. quick one — when the three models (Gemini, Codex, Claude) all run on the same question, how do you handle disagreements, and which one actually gets the final say on the synthesized answer?
@ahmet1wdx "Good question — right now there's no real 'disagreement resolution' happening, because it's sequential, not a debate. Gemini drafts first, Codex does a technical precision pass on top of that draft — Codex doesn't independently answer and then get compared against Gemini, it builds on what Gemini already said.
Claude is meant to be the final synthesis step with actual final say — reviewing both, correcting errors, resolving any inconsistencies — but that step is still being finalized on my end, so responses right now reflect Gemini + Codex without that final arbitration layer yet.
Honestly, once Claude's live, that's exactly its job: not just combining outputs, but actively catching contradictions between the first two and deciding what's actually correct. That's the piece I'm most excited to ship."
How does the offline tutoring mode actually handle updates to the syllabus or new textbook content, especially since CBSE keeps tweaking things every year?
@ngunendi13049 The contents itself isn't automatically detecting CBSE changes — I still need to update it manually when the syllabus shifts. But the DISTRIBUTION is automatic: whenever a phone has internet, it silently syncs any new or updated topics from the server in the background, so a student never needs to redownload the app or manually update anything.
Being fully honest about scale right now: my offline content library is still small — a couple of subjects, not the full Class 6-12 syllabus yet. That's actively being expanded, and the sync mechanism means whatever I add becomes available to every existing user automatically, without needing a new app version.
So: automatic distribution, manual content curation on my end. The alternative — trying to auto-detect syllabus changes — would be a much bigger, riskier problem (accuracy matters a lot for exam prep), so I'd rather be slower and correct than fast and wrong."
How does the three-model pipeline actually handle disagreements when Gemini, Codex, and Claude give conflicting answers, and is there any way for the user to see which model contributed what to the final response?
@bildirci44068 "Honest answer: it's not really 'disagreement resolution' in the sense of comparing independent answers — it's sequential. Gemini drafts first, then Codex builds ON TOP of that draft with a technical precision pass, rather than answering independently and getting reconciled. Claude was designed to be the final arbiter — reviewing both and resolving any real conflicts — but that step is still being finalized on my end.
And yes — you can see exactly which model did what. Every response shows a visible trail (like 'Gemini — first draft → Codex — technical pass → Claude — synthesized final answer'), so nothing happens invisibly. Once Claude's fully live, that trail will show genuine conflict-resolution happening, not just sequential building."
How do you actually handle conflicting outputs when Gemini and Claude disagree on something, and is the offline tutoring side fully on-device or does it still need a sync at some point?
@kayraupf5 "Same honest answer on the conflict side — it's sequential building right now, not independent-then-reconciled (see my reply above for the fuller picture).
On offline tutoring specifically: the on-device model (Gemma 2B) runs genuinely offline — no internet needed for the AI itself once it's downloaded. But there IS a background sync that happens when you DO have internet, pulling updated topics from the server so the app's knowledge stays current over time. So: fully on-device for actually answering questions, with an optional periodic sync to grow smarter — not a hard dependency you need every time you use it."
How does the multi-model setup actually pick which model leads the drafting versus the precision pass on a given query, and can you override that flow when you want one specific model to handle the whole thing?
@nazlcan1332465 "Honest answer: it's not dynamic at all right now — it's a fixed sequence. Gemini always drafts first, Codex always does the technical pass second, regardless of the query type. No override exists yet to force one specific model to handle the whole thing solo.
That's actually a good idea worth building — a 'just use Claude' or 'just use Codex' toggle for when someone already knows which model fits their need best. Adding it to my list."