What is Heys‘ T

Heys'T is an AI-heavy social media application that deeply integrates users and artificial intelligence into the user experience. Its core features range from personalized content delivery and virtual pet companions to AI-powered image generation and an advertising creative studio. The project was developed end-to-end by a single developer, who is me. With design, art direction, and front-end/back-end implementation consolidated into a single workflow, assisted by AI coding tools.

Concept and Framework

Heys'T was conceived as an AI-heavy social media application in which AI functions are fused into nearly every layer of the user experience rather than being attached as an optional add-on. The overall production framework adopted for the project was simple, the developer would personally design the UI/UX and define Heys'T's art style in Figma, then use AI-assisted ("vibe-coding") workflows to implement the application through publication.

The AI matrix selected for Heys'T comprises the following models. Language models: Kimi K2.6, Kimi K2.5, and Moonshot v1-8k. Image generation: Seedream 4.5, accessed via the Replicate API. Video generation: Seedance 2.0, accessed via the Replicate API. Background removal: Bria Remove Background, and Bria Video Remove Background, also accessed via Replicate.

The application was developed using Expo React Native. Expo is built on top of React Native and provides more advanced frameworks and tooling that significantly simplify the development, build, and deployment of cross-platform mobile applications.

The backend was hosted on Supabase, an easy-to-use backend-as-a-service platform that is largely free at low usage levels. Supabase provides authentication, database, storage, and edge function services in a single integrated offering, and its support for Expo is noticeably better than that of Firebase, which made it the preferred choice for this project.

Replicate was adopted as a single point of access to multiple third-party generative models. Because Replicate allows several models to be used through a single API key and consolidated billing, it considerably simplifies payment management and integration compared to maintaining separate accounts with each model provider.

Function Building

The authentication function of Heys'T is built on Supabase Auth. The Supabase plugin available in Cursor allowed the AI agents to read and write most of the framework directly, which significantly accelerated development. Some setup nevertheless had to be performed by hand, notably setting edge function secrets in Supabase so that online functions could be called without storing API keys locally, and provisioning a production-grade SMTP service.

After registration, the user is guided through creating a username and, more importantly, their "Bestie". The user describes the personality and the appearance of their imagined Bestie in free text, and these prompts are then sent to the backend for processing.

The personality prompt becomes the system prompt of the Bestie inside Bestie Chat, the function that lets users converse with their Bestie through the Kimi K2.6 API. Because K2.6 is multimodal with a native web search function, users may send images and videos into the conversation. When search mode is enabled, the request is split across three phases, each implemented as a separate call to Kimi K2.6. Phase 1, understanding. Kimi K2.6 processes the user's request and any media content with reference to the recent conversation history, producing conversation summaries, search summaries, and search request specifications in a structured format passed to Phase 2. Phase 2, web search. Kimi K2.6's web search function retrieves online content based on the Phase 1 specification, and the results are summarised into research notes for Phase 3. Phase 3, response generation. Kimi K2.6 is invoked with the Bestie system prompt and integrates the Phase 2 research notes, the user's current message, and the prior conversation summaries into a long Markdown document for display.

This three-phase pipeline exists for two reasons. First, the Kimi web search function must be called in non-thinking mode, which means a single-pass call cannot reliably reason about what the user actually wants or what the desired output should look like. Splitting comprehension, search, and synthesis into separate calls compensates for this limitation. Second, Supabase Edge Functions enforce a wall-clock limit of approximately 150 seconds per invocation; a single end-to-end call would frequently exceed this threshold and be killed automatically, whereas three shorter calls fit comfortably inside the limit.

The appearance prompt is first analysed by Kimi K2.6, which produces two specialised downstream prompts: one optimised for Seedream 4.5 and one optimised for Seedance 2.0. The Seedream-targeted prompt is sent to Seedream 4.5 via Replicate (called from a Supabase edge function) to generate the Bestie character art, which is then passed through Bria Remove Background to produce a transparent PNG.

The resulting character.png is then used as the first-frame and last-frame reference for Seedance 2.0, together with the Seedance-targeted prompt, to generate an idle animation. The output video is passed through Bria Video Remove Background to produce a transparent video. The final character art and idle animation are rendered on the Bestie page (the file is still named pet.tsx in the codebase, reflecting an earlier naming convention).

After initial creation, the appearance and idle animation can be regenerated from the Bestie settings page. In addition, users can define touch points on the Bestie and create associated touch-point animations to make the character react to taps. Touch-point animations follow the same generation pipeline as the idle animation, with an additional user-customisable prompt.

Users sometimes encounter new internet trends or expressions whose meaning is unclear. The Post Explanation function addresses this case. By long-pressing a post and tapping the explain button, the post's text content is sent to Kimi K2.6 and processed across three phases, again to stay within the 150-second wall-clock limit on Edge Functions. Phase 1. Kimi K2.6 reads the post and extracts possible key terms. Phase 2. Kimi K2.6 performs a web search on the key terms produced in Phase 1. Phase 3. Kimi K2.6 analyses the search results and produces an explanation of the post, rendered under the user's system prompt.

To support creators on the platform, an Ad Studio function was developed. Ad Studio is an AI-driven marketing and content-production workflow in Heys'T that uses campaigns as its organising unit, connecting multi-platform copywriting, poster generation, and feed posting in a single end-to-end path.

When the user creates a campaign, they provide a description of what they want to post or promote. A text generator powered by Kimi K2.6 then produces four platform-specific copy variants tailored to X, Facebook, Instagram, and Bluesky. An optional poster extension lets the user generate accompanying posters with their own prompt through Seedream 4.5. Finally, all generated assets can be passed directly into the Heys'T post-creation flow, allowing the campaign to be published to the feed without leaving the studio.

A Supabase edge function named sync-trends runs on a daily schedule and captures current internet trends. The flow is as follows: the edge function invokes Kimi K2.6 in web-search mode to identify currently trending topics, and then invokes Kimi K2.6 again to perform deeper research on selected topics and turn the findings into trend posts.

The pipeline includes a hard safety limit of a maximum of six rounds per execution to prevent the tool loop from getting stuck or exceeding cost expectations; exceeding this limit is treated as an error. Each day, Kimi searches for candidate topics and selects three of them for detailed research and post generation. Earlier iterations produced more than three posts per day, but since Kimi API usage is billed directly to the developer, three was determined to be a reasonable balance between freshness and cost.

To support more personalized content delivery, Heys'T includes a background function known as Memory Tags. The function runs on a weekly cycle, analyzing each user's recent posting and sharing behavior and passing the collected content to Kimi K2.5 for processing. Kimi K2.5 then condenses these activities into a compact set of keywords that represent the user's recent interests.

The resulting tags serve two purposes within the application. First, they are used by the feed-ranking pipeline to surface posts that share the same tags, increasing the likelihood that the user encounters content aligned with their recent interests. Second, the tags are passed into the user's Bestie context, so that the Bestie is implicitly aware of what its owner has recently been engaging with and can respond to conversations with a more current sense of the user's preferences. Kimi K2.5 was selected for this task rather than K2.6 because the workload is text only. K2.5's lower per-token cost makes it the more economical choice for a recurring batch job of this kind.

Conclusion

Heys'T demonstrates that a single developer, working with a carefully chosen AI model matrix and a vibe-coding workflow, can build and ship a feature-rich AI-driven social media application within a constrained budget. The key decisions that made this possible were the selection of Kimi's model family as the conversational and search backbone, the use of Replicate as a unified gateway to image and video generation models, the adoption of Expo React Native plus Supabase as a low-friction full-stack foundation, and the choice of Cursor with Composer 2 as the primary development environment.