majid@dev ~ $ whoami
Majid Nasiri — Senior Applied AI / ML Engineer
majid@dev ~ $ status
● available Melbourne, Australia · AEST
majid@dev ~ $ deploy --env production --focus llm,rag,agentic,cv,mlops_
10+ years building production AI systems — LLM applications, RAG pipelines, agentic AI workflows, and computer vision, shipped with FastAPI, Docker, and CI/CD.
Senior Applied AI / ML Engineer with a decade of experience building production AI systems across LLM applications, RAG pipelines, agentic AI workflows, and computer vision. Hands-on with OpenAI-compatible APIs, Claude, MCP, LangChain, FAISS/vector retrieval, AI evaluation and observability, FastAPI, Docker, and CI/CD — with a strong focus on practical, scalable, production-ready AI.
production_ready
Ship LLM and CV systems that hold up under real traffic, not just demos.
ground_the_model
RAG and tool-use grounded in real data — FAISS retrieval, schema-aware prompting.
automate_the_pipeline
Reproducible training/export pipelines, Dockerized, wired into CI/CD by default.
Fine-tuned Gemma-4-E2B-it with Unsloth + LoRA on deterministic synthetic datasets for
generating and solving fixed- and variable-size ASCII mazes with randomized start/end positions.
Full pipeline: dataset generation, training, structural validation, merged/GGUF export, llama.cpp
deployment, and an interactive 3D gallery for browsing results.
JetSmartCam is a real-time scene-talker for NVIDIA Jetson Nano that turns live camera frames into detailed scene descriptions. It combines BLIP image captioning and OpenAI-powered visual reasoning with a text-to-speech output pipeline for hands-free awareness of objects, people, actions, and context.
A Jetson Nano voice-assistant pipeline combining WebRTC VAD, Vosk speech-to-text, OpenAI LLM calls, and offline/online text-to-speech to turn live microphone input into conversational responses. End-to-end audio loop, env-based config, Dockerized deployment, and component tests for VAD/STT/LLM/TTS.
An end-to-end maze-solving platform pitting a tool-using LLM agent against a D* Lite baseline under identical fog-of-war constraints for a fair comparison. FastAPI backend, WebSocket streaming solve loop, browser-based live visualization, and provider-agnostic OpenAI-compatible LLM support.
An LLM-powered MCP client that discovers tools from local MCP servers and routes OpenAI function calls to the right backend over stdio.
A terminal tool that prepares an entire codebase—or a selected slice—for LLMs by cleaning,
compressing, and converting it into prompt-ready text. Gitex respects .gitignore, offers
an interactive file picker, extracts Python docstrings and signatures, maps dependencies, inheritance,
and function calls, and can copy the result directly to the clipboard.
Senior AI / ML Developer @ Asset Vision, Melbourne
AVGPT text-to-SQL over asset databases (LangChain, FAISS RAG, Qwen3-30B-A3B on Ollama) · Autopilot multi-model CV inspection microservice (YOLO defect/segmentation/asset detection + monocular depth) · Rewind image matching (ORB, SuperPoint, LightGlue) · Geometry vanishing-point/top-view API · Linear Asset Extraction (SegFormer, SLAM, Kalman/DeepSort) · Pavement Defect Detection (EfficientDet, YOLOv5-9). All shipped via Docker/FastAPI with GitLab CI & GitHub Actions.
Generative AI Engineer @ Freelance, Melbourne
Gemma-4 ASCII-maze generation & solving finetune · JetVoice Jetson Nano voice assistant · MazeLLM tool-using agent vs. D* Lite benchmark · MCP client routing OpenAI function calls to local MCP servers over stdio. See Projects below.
Computer Vision Engineer @ Tosan Techno, Tehran
Face recognition attendance system (MXNet, OpenCV, Dockerized FastAPI) · Face spoofing / anti-biometric-attack detection (TensorFlow, PyTorch, OpenCV).
Computer Vision Developer @ Digibina, Tehran
AR virtual try-on platform for glasses/sunglasses — facial landmark detection, glasses estimation, and 3D fitting (OpenCV, DNNs, Python/C++), deployed as FastAPI microservices on Docker.
ML / CV Developer @ Freelance / Contract, Tehran
SegFormer-based road-line segmentation & classification (PyTorch, C++) · Capsule-network 3D point-cloud classification (PointNet, PointConv, DGCNN) and zero-shot 3D classification with BERT + contrastive learning.
languages:
- python
- c++
- bash
llm_and_agents:
- openai-api / claude / mcp
- langchain
- agentic workflows
- rag (faiss, vector retrieval)
- prompt engineering
- qwen3-30b-a3b (ollama / vllm)
computer_vision:
- yolo (v5-v9, v26x) / fastsam
- segformer / efficientdet
- pytorch / tensorflow / libtorch / torchscript
- mxnet
classical_cv_and_tracking:
- orb / sift / surf / superpoint / lightglue
- slam / kalman filter / deepsort
infra:
- docker / docker-compose
- fastapi
- ci/cd (gitlab ci, github actions)
- ai evaluation and observabilityThesis: Semantic Image Segmentation in outdoor scenes using Deep Neural Networks.