Real engineering stories: what we broke, what we learned, and how we built an AI meeting assistant from scratch.
Every zero-cost AI model on OpenRouter, Groq, Google, and beyond — with rate limits, context lengths, and best use cases compared. No credit card required.
Our meeting summary fired at 30s, then 5 minutes, then never. The deep dive into JavaScript timer mechanics — and the recursive setTimeout pattern that fixed it permanently.
The xAI Grok integration: the architecture decision, the day-one 400 error from a wrong model name, and what dual-AI actually unlocks for real-time correction latency.
We built speaker diarization, shipped it, realized the Web Speech API makes it fundamentally impossible to do well, then deleted everything. The full post-mortem.
VORA started as Python FastAPI + Faster-Whisper. We deleted all of it. A 3:1 fix-to-feature ratio forced the question: what if we just used the Web Speech API?
Benchmarks, the CORS/SharedArrayBuffer nightmare, and honest numbers on when browser Whisper is viable for real-time Korean speech recognition — and when it just isn't.
Three complete redesigns: parallel (race conditions), mutex (priority inversion), and finally the priority queue. Each failure explained in detail with real code.
Three UI overhauls. Two weeks arguing about an equalizer animation. The design decisions that transformed VORA from "functional prototype" to "product people compliment."
The Web Speech API returns multiple hypotheses. Our domain-aware local reranker selects the best one — achieving 8% WER improvement with zero API calls using technical term scoring.
Three iterations: text dump, template HTML, AI-generated 개조식 minutes. Prompt engineering to make Gemini write proper Korean business documentation, plus a custom markdown table renderer.
Dual HTML files, hreflang tags, the language consistency audit that revealed 12 broken links, and why we chose file duplication over i18n libraries — with honest hindsight.
Three name changes, a "Voice Oracle" detour, and why renaming a deployed product is much more expensive than picking the right name at the start.
The sitemap/robots.txt placement that cost us 3 weeks of Google indexing, CORS header conflicts with Google Fonts, and the _headers file format every Cloudflare Pages developer needs.
Building real-time AI speech correction nearly killed the Gemini API. Here's how we built a 4-layer defense architecture that eliminated 429 errors completely while maintaining sub-second perceived latency.
Alibaba's Qwen3-ASR is a game-changer, addressing Whisper's limitations. From the lightweight 0.6B model to singing recognition and emotion analysis, find out why this model is the new standard.
In February 2026, which Gemini model is the fastest and smartest? From TPS/TTFT scorecards to a surprising twist in coding performance, here is your guide.
The VORA team is conducting internal tests on Alibaba's high-performance lightweight model, SenseVoice Small. Exploring technical features and on-device feasibility.
Reviewing Whisper v3-large as VORA's next-gen engine. Testing feasibility in browser environments and Korean recognition performance.
Google released Gemini 2.0 Flash on Feb 5, 2025. Summarizing the core of its 2x faster processing and multimodal capabilities.