Applied AI Intern

Job Description Duration: 6 Months Location: IIT Madras Research Park, Chennai We are seeking a highly driven Applied AI Intern to help build a multi modal evaluation engine. You will take state-of-the-art, open‑weight Vision‑Language Models (VLMs) and adapt them to solve a highly specific, complex problem: reading dense regulatory standards (like SOC 2 or ISO 27001) and evaluating user‑submitted evidence—ranging from AWS configuration screenshots to PDF audit reports—to determine compliance. This is an engineering‑heavy internship focused on data pipelines, Retrieval‑Augmented Generation (RAG), and parameter‑efficient fine‑tuning, with zero tolerance for model hallucination. Phase 1: Data Infrastructure & Baselines (Months 1‑2) Construct a robust “Golden Dataset” of highly curated GRC evidence. Deploy baseline open‑source VLMs (e.g., LLaVA‑NeXT, Pixtral) to establish initial accuracy metrics. Overcoming data scarcity; ensuring absolute accuracy in the ground‑truth labels alongside domain experts. Phase 2: Multimodal RAG & Fine‑Tuning (Months 3‑4) Implement a retrieval pipeline to feed relevant compliance clauses to the model at inference. Apply Parameter‑Efficient Fine‑Tuning (PEFT/LoRA) to adapt the VLM to complex UI dashboards and terminal outputs. Prevent catastrophic forgetting; optimizing retrieval chunking strategies for dense legal/technical text. Phase 3: Red Teaming & Evaluation (Month 5) Develop an automated evaluation framework to rigorously test for hallucination rates, edge‑case failures (e.g., dark mode interfaces, low‑res screenshots), and false positives. Aggressive adversarial testing; forcing the model to fail so vulnerabilities can be patched. Phase 4: Deployment & Handoff (Month 6) Wrap the tuned model and RAG pipeline in a clean, production‑ready API (FastAPI). Document the entire data and training pipeline for internal teams. Optimize inference latency; ensuring code maintainability. Core Responsibilities • Design and execute data engineering pipelines to extract, sanitize, and format complex cybersecurity frameworks and visual evidence into machine‑readable datasets. • Implement and optimize Multimodal RAG architectures using vector databases and embedding models. • Fine‑tune open‑source large language and vision models using Hugging Face tools, specifically targeting edge cases where off‑the‑shelf models fail to read specific UI elements. • Work directly with internal domain experts to validate model outputs and aggressively iterate on failure cases. Requirements • Academic Background: Pursuing a Master's or PhD in Computer Science, Applied Mathematics, or a related technical field. • Applied ML Engineering: Strong proficiency in Python and PyTorch. You build reliable systems, not just theoretical proofs of concept. • Ecosystem Mastery: Deep, hands‑on experience with the Hugging Face ecosystem (Transformers, Datasets, PEFT, Accelerate). • RAG & Vector DBs: Proven experience building retrieval systems (using tools like LangChain, LlamaIndex, Pinecone, or Chroma) and understanding chunking strategies for dense text. • Analytical Rigor: A highly critical approach to model evaluation, with an understanding that in GRC, “mostly correct” is a failure state. Benefits & Terms of Engagement • Salary - As per entry level role standards for the duration of internship payable monthly. • Location - Crossbow Office in IIT Madras Research Park • Workhours - Flexible as per existing policies with Crossbow, driven by meeting project objectives. Remote work can be evaluated as the project progresses.