atypica.AI Changelog

Previously, AI research jumped straight into interviews and search. Now every study begins with a visible plan that outlines who will be contacted, what will be investigated, and why each step matters.

The planning step automatically selects the right business analysis frameworks (JTBD, STP, GE matrix, etc.) so teams understand the strategic lens behind the upcoming interviews and desk research.

Before the final report is generated, the AI prepares a detailed research log that summarizes collected information and key decisions to give context behind the conclusions.

🚀 Full Launch: Interview Project & AI Persona Import

Both features now available to all plans (removed preview/beta status)

Free/Pro plans: Access to limited features
Max/Team plans: Access to full functionality

✨ Major Feature: AI Interview Question Optimization

AI automatically analyzes your interview project brief and optimizes questions for better research outcomes

Transforms complex questions into clear, answerable segments following research best practices

Applies fact-first methodology and generates granular questions to capture deeper insights

Works seamlessly with both human interviews and AI persona interviews

🎨 UX Improvements

Enhanced voice input speed and responsiveness

Redesigned persona import workflow with better status visibility

Added team workspaces, supporting member invitations, removal, and role management. Implemented user identity switching for seamless work between personal and team spaces. Introduced dedicated routes and navigation menus for the team edition.

Launched the core AI Persona Import feature, opening up a public beta for Max plan users. Users can upload their own interview transcripts (PDF, JSON, CSV) to build completely private Human AI Personas (Proprietary). The AI provides multi-dimensional analysis and scoring for the uploaded content, generating an interactive AI Persona and visualizing the analysis with tools like radar charts.

Integrated Groq Whisper API for near real-time, streaming transcription, dramatically improving the fluency of voice interviews (2025-07-04).

Completed the core development of the new Interview Project and moved it into the internal Alpha testing phase. The new workflow aims to simplify the creation and sharing process for interviews (2025-07-24).

Added a dedicated attachments table to manage user-uploaded files centrally, enabling file reuse and optimizing storage (2025-07-20).

Completed the core database migration from China to the US region, laying the foundation for global services (2025-06-29). Launched the new new-study-interview feature, clarifying user research goals through an initial AI conversation. Published the first global launch announcement on Twitter, officially marking the product's entry into the international market.

Completed the core development of the Product R&D research flow and moved it into the internal Alpha testing phase (2025-06-27).

Launched a completely redesigned official website (atypica.ai) with an improved information architecture and visual experience, featuring all AI-generated illustrations (2025-06-27).

Launched the MAX subscription plan, offering more tokens and advanced features (2025-06-12). Implemented plan upgrades/downgrades with automated pro-rated pricing based on remaining tokens (2025-06-26). Transitioned to USD-based pricing and integrated Stripe's Alipay auto-debit feature to streamline payments (2025-06-19).

Integrated Tavily for real-time web searches, providing richer context during the initial research phase. Added Twitter as a new data source for building more globally relevant user personas. Introduced the "Real Persona" concept, allowing personas based on actual interviews to be highlighted during research.

Implemented automatic classification of research queries (Testing, Planning, Insights, Creation) to match them with distinct research flows and report structures (2025-06-09).

Integrated Google Imagen 4.0 and Midjourney to dynamically generate high-quality, context-aware images directly within research reports (2025-05-25).

Enabled prompt caching for Bedrock Claude. By hashing historical messages, it reduces token consumption for repetitive requests by 90% (2025-05-29).

Optimized the voice recognition feature using the Groq Whisper model for instant speech-to-text conversion, significantly improving interaction speed. Fixed a bug where prompt caching failed due to an issue in the Vercel AI SDK.

Users can now upload documents (PDF, etc.) as context when initiating a study. The AI references this information during subsequent interviews and report generation (2025-05-18).

Integrated the pgvector extension into the database, adding Embedding vectors to Personas to enable semantic, similarity-based searches (2025-05-13).

Deployed a standalone html-to-pdf service, allowing research reports to be downloaded as high-quality PDF files with a single click (2025-05-12).

The Study Agent can now pass more specific instructions and context to the Interview Agent, receiving richer feedback and significantly reducing overall token consumption.

Successfully migrated the backend database from MySQL to PostgreSQL, paving the way for advanced features like vector search and complex queries (2025-04-29). Resolved initial performance issues post-migration by adding indexes to nerd stats queries.

Introduced the gpt-4o model with support for parallel tool calling to effectively reduce research costs (2025-04-16). Adopted Google Gemini 1.5 Flash for specific tasks to further optimize costs (2025-04-21). Implemented strict token and step limits to prevent research tasks from running over budget or time (2025-04-28).

Switched the billing system from simple "points" to a more granular Tokens-based model. Officially launched the Pro subscription plan and a new Pricing page (2025-04-20). Integrated Stripe to support global credit card payments (2025-04-09).

Split ChatMessage data into a separate table, resolving data override issues during concurrent writes and improving system stability and scalability (2025-04-12).

Added TikTok and Instagram as new data sources to broaden consumer insights (2025-04-17).

Introduced a user points system and launched the first payment feature ("buy a coffee") via the Heidian API. Added a Hello Agent to capture enterprise leads through conversation.

Added an invitation code system for early user onboarding (2025-03-31).

Overhauled the multi-step process into a unified "research-driven chat" interface, where users can kick off a full study just by asking a question. Introduced the Study Agent as a "commander" to plan tasks, orchestrate specialized agents, and report back to the user.

Adopted a split-screen design (left: chat, right: atypica llm console) to provide real-time transparency into the AI's thought and execution process (2025-03-19). Implemented a "Replay" feature with shareable links (2025-03-22). Added "Nerd Stats" for tracking token and step consumption (2025-03-26). Added a light/dark theme toggle (2025-03-25). Generated the first changelog page automatically from git log.

Launched a multi-page, guided scout → personas → analyst → interview → report process, establishing a foundational framework for market research.

Analyst: Create and manage research topics, stored in the database.
Interview: Simulate AI-driven interviews with Personas.
Report: Automatically generate HTML reports with a shareable link.

Introduced user authentication (login/signup) and made research history private (2025-03-16).
First deployment to an AWS cluster at atypica.musedam.cc, with GitHub CI for continuous integration.

Product concept defined: build interactive User Personas from social media data to test product innovation ideas (2025-03-05).

Achieved the first end-to-end flow: automatically summarizing and creating User Personas from "Xiaohongshu" data. The SavePersona tool successfully saved the first AI-generated persona to the database.

Next.js + Vercel AI SDK + LLM Tools as the core framework.
Claude 3.7 Sonnet as the primary reasoning model.
Significantly reduced early token costs by rewriting prompts with an experimental method (2025-03-08).