Artificial Intelligence Intern

AI Engineering Intern (6 Months) | LLMs, Edge AI, Agentic AI, Documentation AIBrainy Neurals | Ahmedabad | Work from office onlyBrainy Neurals is a founder-led AI engineering company based in Ahmedabad. For nine years we have shipped production AI for clients across 15+ industries, from vision models running on NVIDIA Jetson boards to agents that draft regulated pharma documents. We are an NVIDIA Inception member and ISO 27001 certified.We are hiring interns who understand what happens inside a model, from the attention layer down to quantized weights running on a chip. Calling an API and writing a decent prompt is useful, and this internship expects much more than that from you.Please read this part twice before you apply. This internship has no data analysis, no dashboards, no Excel or Power BI reporting and no Kaggle-style tabular modelling. If data science is the career you want, the role will frustrate you within a month, so save your application for a better fit.What you will work onLLM engineering: fine-tuning open-weight models with LoRA and QLoRA, and building evaluation sets that show whether the fine-tune helped.Edge AI: quantizing and converting models with TensorRT and ONNX, then measuring latency and accuracy on NVIDIA Jetson hardware.Agentic AI: building agents that plan, call tools, recover from failed steps and leave a trace you can debug.Documentation AI: parsing PDFs and scanned pages, and generating GMP documents for pharma that a QA reviewer can sign off.The basics we will test you onYou should be able to explain each of these at a whiteboard, without notes, in plain words. The technical interview is built on this list.How self-attention works, and why its compute cost grows with the square of the sequence length.The difference between a decoder-only model like Llama and an encoder-only model like BERT, and which jobs suit each.When you would pick full fine-tuning, LoRA, QLoRA or DPO over prompting and retrieval, with a reason for each choice.What the KV cache, context length, batch size and INT8 quantization do to latency and memory on a GPU with limited VRAM.How one training step runs end to end, from the forward pass and loss through backpropagation to the optimizer update, and how you read overfitting from the loss curves.Must haveA B.E., B.Tech, M.E. or M.Tech in CS, IT, EC or AI/ML, completed in 2026 or due in 2027.Strong Python and hands-on PyTorch, including at least one training loop you wrote from scratch.At least one project where you trained or fine-tuned a model and measured the result, with code on GitHub you can explain line by line.The ability to read a research paper or a model card and turn it into working code.Good to haveHands-on use of Hugging Face Transformers, vLLM, llama.cpp or Ollama.Time spent deploying a model with TensorRT or ONNX Runtime, on a Jetson board or any edge device.An agent you built, with LangGraph or plain Python, that finishes a real multi-step task.Work on OCR or PDF parsing where tables and scanned pages had to come out clean.How the internship worksSix months, full time, in our Ahmedabad office. We offer no remote or hybrid option at any stage of the internship.You will work on live client projects and our own products, with a senior engineer reviewing every pull request.You will receive a monthly stipend of [amount].Final-year students need a letter from their college confirming they can attend full time for all six months.How to applySend an email to humans@brainyneurals.ai with the subject line “AI Intern | Your Full Name” and include these four things:Your resume as a PDF.Your GitHub profile link, plus the one repository you want us to read first.Your answers to any two of the questions below.The earliest date you can join full time.Pick any two of these questions:Your 7B model answers well in FP16 and badly after INT8 quantization. What do you check first, and why?A client hands you 300 internal SOPs and wants a model that answers questions from them. Would you fine-tune or build retrieval, and how would you prove your choice was right?Your agent keeps calling the same tool in a loop and never finishes the task. How do you find the cause and stop it from happening again?Training loss keeps falling while validation loss climbs after epoch three. What is going on, and which two changes would you try first?A scanned 40-page GMP document has tables that break across pages. How would you extract them so every row stays intact?Write these answers yourself, under 150 words each. In the technical interview we will ask you to defend every line at a whiteboard, so answers pasted from ChatGPT or Claude will cost you the interview. Applications that arrive without two answers will not be read.How selection worksOur engineering team reads your two answers and your GitHub repository.Shortlisted candidates come to our Ahmedabad office for a technical interview on the basics listed above.You then complete a hands-on task in the office, debugging or extending real model code.The final round is a conversation with our leadership team about the work you want to do here.