Muhammad Haris B.

Top Rated AI engineer, 100% Job Success, 14 completed projects. I fine-tune LLMs and image models (LoRA/QLoRA), build RAG pipelines and AI agents in Python with LangChain, OpenAI and Claude, and ship them as monitored APIs on AWS or GCP. Most clients come to me with one expensive bottleneck. AI is the tool; the bottleneck is the target. RECENT RESULTS - Fine-tuned a Flux image model for a B2B client ($3,200 project, 5-star, "went above and beyond") - Trained 30 production LoRAs on a custom base model for a generative image product - Built a multi-modal AI video generation system (script to voice to video) for a fitness client - Delivered a Python ML backend for an NGO's image-recognition API - Wrote and now maintain an automation bot under a rolling retainer (12+ months, 5-star) WHAT I BUILD LLM Fine-Tuning: LoRA, QLoRA, RLHF, quantization and domain adaptation with PyTorch and Hugging Face. OpenAI GPT-4, Anthropic Claude and Gemini via API, or open-weight models (Llama 3, Mistral, Qwen) served through vLLM or Ollama when privacy or cost per token favors self-hosting. RAG Systems: Retrieval augmented generation over your documents with Pinecone, Weaviate or pgvector, orchestrated in LangChain or LlamaIndex and exposed through FastAPI. Embedding selection and chunking are where RAG succeeds or fails, so I evaluate retrieval accuracy before I touch the prompt. AI Agents and Automation: Multi-agent workflows with LangGraph and MCP, plus n8n, Make and Zapier for lighter automation, connected to your CRM, Slack or Google Workspace. AI Chatbots and Voice Agents: Customer support, lead qualification and internal assistants on web, WhatsApp and voice, with Whisper speech-to-text and guardrails wherever a wrong answer is expensive. Generative Image and Computer Vision: Flux and Stable Diffusion fine-tuning, LoRA training, object detection, OCR and document extraction with YOLO, OpenCV and PyTorch. Deployment: Docker on AWS, Azure or Google Cloud with logging, rollback and evaluation pipelines that track hallucination rate, latency and cost as three separate numbers. HOW I WORK A model that sounds confident and is wrong does more damage than one that admits it doesn't know. I ask more questions upfront than most freelancers, because a model trained on a misunderstood problem is expensive to unwind later. Send me the bottleneck you've stopped counting and I'll tell you honestly whether AI is the right tool or whether something simpler would do it.