@Hamna

here

Changed the UI a bit

There were some UI erros in the git pages file i made, so this is the last final verison for this project for the demo link: hope you’ll like this : )

Fixed it today! I replaced ChatInterface in the Space with a gr.Interface that takes the history as a JSON input, so the conversation actually reaches the model now. I also added a history count to the logs so I can see it arriving with every message. This fixes Amadeus too, since they share the same Space. Iris finally remembers what we’re talking about, and that feels like the right place to call this project complete

Spent today debugging the memory problem. The frontend was sending the chat history and the router was clearly using it, but the model was answering like every message was the first one. I asked “what did I just ask?” and got a totally random reply. Then I checked the Space’s API info and found the problem: Gradio 6’s ChatInterface hides the history from the API, so it was silently throwing the whole conversation away

Tested the new pipeline today. I asked “who won the last cricket world cup?” and it searched, found the facts and answered “Australia” correctly. Iris knowing current events feels like a huge upgrade. But while chatting normally I noticed something weird: it kept saying “Hai haii” to everything and had no idea what we’d been talking about

Big backend day. I made Iris use the exact same backend as my other project, Amadeus: the same Hugging Face ZeroGPU Space, plus a RAG pipeline. Now a small router model on Hack Club AI rewrites the question, decides whether it needs a web search, and remembers facts about the user. If it needs current info, it searches the web with Exa and passes verified facts to the model. The UI stayed exactly the same

Tried Modal as another hosting option today, to compare it with ZeroGPU. I switched the route to call a Modal endpoint with an API key. It worked, but managing two setups for two projects (Iris and Amadeus) was getting messy, and I started thinking about putting both on the same backend

Came back to Iris after a while because responses were way too slow on free CPU hosting. Today I moved the model to Hugging Face ZeroGPU, so it actually runs on a GPU now. I rewrote the backend route to call the Space through the Gradio API, with retries and fallback messages in case something fail

Final UI touches today. I polished the About modal, updated the project links and rewrote the README with the whole training pipeline. Then I deployed it on Vercel. Iris is officially live and I shared the link with everyone

Today I connected the frontend to the real Hugging Face backend instead of the placeholder. Seeing Iris reply on my own website, in Senpai’s style, felt unreal. Responses were slow, but it was talking, and I spent way too long just chatting with it

I replaced the old playlist with songs that match the vibe better and made new covers for each track. Some of the song links broke right after, so I fixed those, and one mp3 was way too big, so I compressed it. Small stuff, but the radio station panel feels complete now

I didn’t like how the UI looked yesterday, so today I redesigned almost all of it. I went with a soft lavender pastel theme that feels more like Iris, removed the cyber cursor and added a cute Senpai avatar that pops up with little messages. I also made the whole thing responsive so it works on phones too

Started the frontend today with Next.js and Tailwind. I built the main layout: a chat panel in the middle, a music player on the left, and a side panel on the right. I also added a loading screen and little send/receive sound effects, because a chatbot without sounds feels lifeless. Pushed the first commit with the UI, routing and model connection all together

Training ran for 1000 steps today and I watched the loss drop from about 2.5 to 0.3. When I tested it, it replied with “Kk” and an eye roll emoji, and that’s literally how Senpai texts. I sent screenshots to my friends and they were all shocked. That was the best moment of the project so far

Guyss!! This is gonna be the last change I’ve made today, Senpai chatbot is live now and she’s talking like my friend (really couldn’t be more happy to see her talk to me again 🥹 )

Made some Minor UI changes on frontend nextjs

Set up my Kaggle notebook today with Unsloth and Llama-3 8B as the base model. Most of the day went into fighting configs and memory errors on the T4 x2 GPUs. After a lot of trial and error I settled on a 2048 sequence length and a 5e-5 learning rate. It finally started training without crashing, and I almost cried