main.py

from fastapi import FastAPI, File, UploadFile, Form
from fastapi.responses import HTMLResponse, JSONResponse
import torch
import torchaudio
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq, AutoTokenizer, AutoModelForCausalLM

app = FastAPI()

# Load IBM Granite models once on startup
SPEECH_MODEL = "ibm-granite/granite-speech-3.3-8b"
LLM_MODEL = "ibm-granite/granite-3.3-8b-instruct"

speech_processor = AutoProcessor.from_pretrained(SPEECH_MODEL)
speech_model = AutoModelForSpeechSeq2Seq.from_pretrained(SPEECH_MODEL).to("cpu")
tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL)
llm_model = AutoModelForCausalLM.from_pretrained(LLM_MODEL).to("cpu")

@app.get("/", response_class=HTMLResponse)
def home():
    return """
    <h1>Meeting Memory Workflow Automation</h1>
    <form action="/transcribe" enctype="multipart/form-data" method="post">
      <input type="file" name="audiofile" accept="audio/*" required/>
      <button type="submit">Upload & Transcribe</button>
    </form>
    """

@app.post("/transcribe")
async def transcribe(audiofile: UploadFile = File(...)):
    audio_bytes = await audiofile.read()
    import io
    wav = io.BytesIO(audio_bytes)
    audio, sample_rate = torchaudio.load(wav)

    inputs = speech_processor(audio, sampling_rate=sample_rate, return_tensors="pt").to(speech_model.device)
    generated_ids = speech_model.generate(**inputs)
    transcript = speech_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

    # Summarize
    prompt = f"Summarize the following text: {transcript}"
    inputs = tokenizer(prompt, return_tensors="pt").to(llm_model.device)
    summary_ids = llm_model.generate(**inputs, max_new_tokens=200)
    summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True).replace(prompt, "").strip()

    html = f"""
    <h2>Transcript</h2>
    <pre>{transcript}</pre>
    <h2>Summary</h2>
    <pre>{summary}</pre>
    <a href="/">Back</a>
    """
    return HTMLResponse(content=html)

# Add any orchestration/agent endpoints as needed!