অ্যাডভান্সড প্রজেক্ট

লক্ষ্য: Planner-Researcher-Writer-Critic agent graph — input topic → cited report (PDF)।

1. Tool spec — web search, scrape, summarize, cite।
2. LangGraph দিয়ে stateful workflow (retry, branch)।
3. Critic agent → quality loop (max 3 iter)।
4. Postgres-এ trace, cost log।
5. Streamlit/Next.js UI + PDF export।

Stretch: Human-in-the-loop approval before publish।

লক্ষ্য: OpenAI-substitute — multi-user, multi-conversation, tool-calling।

1. vLLM serve (4×A10G বা 1×A100) — OpenAI-compatible API।
2. Function calling — calculator, web search, code exec।
3. RAG — user document upload → per-user collection।
4. Auth, billing (Stripe), rate limit।
5. Token streaming, message edit, export।

Stretch: Voice mode (Whisper + TTS)।

লক্ষ্য: X-ray/CT-তে pathology detect + Grad-CAM heatmap + radiology report।

1. Public dataset (ChestX-ray14, RSNA) preprocess।
2. Swin-Transformer fine-tune — multi-label classification।
3. Grad-CAM/Attention rollout heatmap overlay।
4. BLIP-2 দিয়ে natural language report generate।
5. DICOM upload UI, HIPAA-style PII redact।

Stretch: Active learning loop — radiologist feedback retrain।

লক্ষ্য: Click stream → online feature update → low-latency personalized rec।

1. Synthetic event generator → Kafka।
2. Flink — sliding window aggregate → Feast online store।
3. Two-tower model (user/item embedding)।
4. TorchServe inference, <50 ms p95।
5. A/B framework — model A vs B, lift report।

Stretch: Bandit-based exploration (LinUCB)।

লক্ষ্য: 100M–350M param Transformer — Bangla+English corpus-এ pretrain।

1. Corpus সংগ্রহ (Wikipedia BN/EN, CulturaX) → 5-10B token।
2. BPE tokenizer train (vocab 32k)।
3. GPT-style decoder — 12 layer, 768 dim, RoPE, FlashAttn।
4. FSDP/DeepSpeed-এ 4-8 GPU training (RunPod/Lambda)।
5. Eval — perplexity, downstream task (sentiment, NLI)।
6. HuggingFace Hub-এ release + model card।

Stretch: Instruction-tune + DPO → chat model।