Alif Tasbir

about

I build machine learning systems end to end, from deep learning research to deployed applications.

I'm a recent MSc Artificial Intelligence graduate from De Montfort University. My work covers deep learning for biomedical signals such as EEG and ECG, using CNNs, LSTMs and graph neural networks. It also covers LLM fine-tuning, retrieval-augmented generation and multi-agent systems, as well as efficient inference on edge hardware. I take models from experiment to product with PyTorch, FastAPI, Docker and MLflow, and I've published three peer-reviewed papers in clinical AI. I'm looking for an AI/ML Engineer role building machine learning solutions for real-world problems.

projects

● live demo MSc dissertation

ECG Multi-Agent Triage: Heterogeneous Edge AI for Clinical Decision Support

A two-agent ECG triage system. A CNN-LSTM perception agent classifies heartbeats and passes events to a frozen Gemma reasoning agent, either as text or as learned virtual tokens. It includes guardrails, an orchestrator and clinician Q&A. This is a research prototype.

• 24–78% faster LLM prompt processing than compact text, with balanced accuracy of 0.83 vs 0.73 at 10 and 20 events.
• 10–30% higher throughput and 12–28% less energy per decision on a 12 GB GPU, measured with NVML at 100 Hz.

Try the live demo ↗ App code Research code

Hosted on Streamlit's free tier. If the app is asleep, click "wake up" and give it about 30 seconds.

PythonPyTorch · Gemma · Streamlit

Fine-Tuned Small LLM for Multi-Hop RAG

• LoRA fine-tuned Qwen3-1.7B: exact match 10.4 → 58.4 on HotpotQA, matching a 4.7× larger Qwen3-8B with ~5× smaller weights at ~8× the throughput.
• Hybrid retrieval (BM25 + FAISS, reranking) with multi-hop search: evidence recall 54% → 89%, with citations and abstention.
• MCP server with typed tools and OpenTelemetry tracing; tool-use fine-tuned LangGraph agent (+19.7 F1).

PythonPyTorch · PEFT · LangGraph · MCP · FAISS

Multimodal RAG Application

• Bilingual (Bangla & English) RAG over PDFs, scans, Office files and images, with Tesseract OCR, answering in the user’s language with page-level citations.
• Hybrid retrieval: BM25 plus multilingual-e5 embeddings in FAISS, merged with reciprocal rank fusion, in per-session document stores.
• LLM-routed web search: the model decides when fresh information is needed, rewrites the query and sizes the Tavily search. Streams answers from Gemini or Groq (gpt-oss-20b) via FastAPI + Streamlit.

PythonFastAPI · Streamlit · FAISS

AI Interior Visualizer

An interior design visualizer built for the Nano Banana hackathon, powered by Gemini 2.5 Flash Image.

TypeScripthackathon

All repositories on GitHub →

experience

education

publications

1. Parkinson's disease detection from EEG signal employing autoencoder and RBFNN-based hybrid deep learning framework utilizing power spectral density
   Digital Health · SAGE · 2024
2. Graph attention network and radial basis function neural network-based hybrid framework for epileptic seizure detection from EEG signal
   Journal of Autonomous Intelligence · Vol. 7, Issue 3 · 2023
3. Epileptic seizure detection from EEG signals using linear graph convolutional network and DenseNet based hybrid framework
   Journal of Radiation Research and Applied Sciences · Elsevier · Vol. 16 · 2023

skills

llms & rag

LoRA/QLoRA, PEFT, TRL, Hugging Face Transformers, LangChain, LangGraph, Model Context Protocol (MCP), Ollama, FAISS, ChromaDB, hybrid search & reranking, Sentence Transformers, Gemini API, LLM evaluation (EM/F1, faithfulness, BLEU/ROUGE/BERTScore)

ml & dl

PyTorch, TensorFlow, Keras, scikit-learn, Pandas, NumPy, CNNs, LSTMs

apis & apps

FastAPI, Flask, Django, Streamlit, Gradio

mlops

Git, GitHub Actions, Docker, MLflow, DVC, OpenTelemetry

languages

Python, C, C++, Java, JavaScript, SQL, HTML, CSS

contact

I'm open to AI/ML Engineer roles and research collaborations, so feel free to get in touch.