I am an M.Sc. student in Artificial Intelligence at Brandenburg University of Technology Cottbus–Senftenberg with 4+ years of professional software engineering experience building scalable backend, cloud-native, and enterprise banking systems.
I currently work as a Student Assistant at BTU, collaborating with a research supervisor to develop and benchmark robust image-registration methods under noise and illumination changes. My AI work spans LLM pretraining and fine-tuning, Generative AI, agentic systems, multimodal learning, computer vision, and MLOps.
My engineering approach emphasizes reproducibility, rigorous evaluation, typed interfaces, fault-tolerant workflows, deployment readiness, and measurable system performance.
Programming: Python, Java, SQL
AI & Frameworks: PyTorch, Transformers, OpenCV, scikit-image, scikit-learn, LangGraph, LangChain
Backend: FastAPI, Spring Boot, REST APIs, microservices
MLOps & DevOps: Docker, Kubernetes, GitHub Actions, MLflow, DVC
Cloud, Data & Tools: AWS (EC2, S3), PostgreSQL, MongoDB, FAISS, Pinecone, Git, Linux, Bash
June 2026 – Present · Cottbus, Germany
- Collaborate with a research supervisor to develop an image-registration pipeline robust to noise and illumination changes.
- Implement and benchmark classical and hybrid methods using accuracy, runtime, CPU usage, and memory consumption.
October 2017 – March 2022 · Dhaka, Bangladesh
- Automated CI/CD and standardized Docker/Kubernetes deployments for enterprise systems.
- Modernized distributed backends using Spring Boot microservices, REST APIs, workflow automation, and PL/SQL tuning.
- Built highly available services and parallel ingestion pipelines for mission-critical data and production operations.
A from-scratch medical language-model project covering architecture design, tokenizer training, general pretraining, medical continual pretraining, supervised fine-tuning, and rigorous evaluation.
- Built a 35.5M-parameter decoder-only Transformer with RoPE, RMSNorm, SwiGLU, causal attention, tied embeddings, and a custom 16K byte-level BPE tokenizer.
- Developed deterministic mixed-precision training with gradient accumulation, atomic checkpoints, SHA-256 lineage, and exact crash recovery on compatible runtimes.
- Pretrained on 239.5M general-domain tokens and continued training on 263.2M packed tokens using a 70/30 medical-to-general rehearsal mix.
- Reduced medical validation perplexity from 32.06 to 23.00 while operating within a predefined general-retention budget.
- Applied response-masked SFT across seven instruction datasets, improving sealed-test response perplexity by 21.4%.
Tech: Python · PyTorch · CUDA · Transformers · NLP
A production-oriented fact-checking platform that coordinates specialized research agents while enforcing evidence quality, cost limits, and publication safety.
- Designed supervisor-agent orchestration for five concurrent specialist agents covering primary, academic, web, fact-check, and contradiction research.
- Implemented typed permissions, immutable assignments, hard cost budgets, evidence-sufficiency routing, human escalation, and fail-closed publication controls.
- Built a durable 12-stage investigation workflow with numerical and temporal verification, defender–challenger reasoning, source-independence analysis, and sentence-level citation assurance.
- Added checkpoints, server-sent event progress, cancellation, and duplicate-charge prevention for reliable long-running execution.
Tech: Python · FastAPI · LangGraph · Pydantic · OpenAI · SerpAPI · Next.js
A multimodal industrial-inspection system that fine-tunes Qwen2.5-VL for structured anomaly analysis across multiple products and task types.
- Fine-tuned Qwen2.5-VL using 4-bit QLoRA with 3.02M trainable parameters.
- Engineered 52,863 leakage-safe instructions from 10,821 images across eight tasks and twelve product categories.
- Achieved 85.0% anomaly accuracy, 99.3% product accuracy, 99.3% report-schema validity, and 98.0% appropriate-abstention accuracy on uncertainty cases.
- Evaluated on a frozen 2,100-record suite with resumable GPU workflows, deterministic evaluation, release gates, model versioning, and rollback.
Tech: Python · PyTorch · Transformers · QLoRA · PEFT · Pydantic · pytest
-
M.Sc. in Artificial Intelligence - Brandenburg University of Technology Cottbus–Senftenberg, Germany
October 2023 – Present -
B.Sc. in Computer Science and Engineering - American International University–Bangladesh
January 2013 – December 2016
- Generative AI and Large Language Models
- Agentic AI and multi-agent systems
- Multimodal learning and Computer Vision
- AI/ML Engineering and MLOps
- Reliable, evaluation-driven AI systems
I am open to AI/ML Engineering, Generative AI, Agentic AI, Computer Vision, MLOps, working-student, internship, and research opportunities where I can combine production software engineering with rigorous AI development.
I am also interested in collaborating on open-source projects involving LLMs, multimodal systems, agentic workflows, evaluation, and deployment.
- LinkedIn: linkedin.com/in/moshiur00
- GitHub: github.com/moshiur00
- Email: moshiur.mishuk@gmail.com
Thank you for visiting my profile. Feel free to connect, collaborate, or explore my projects.