nandhu@ai-systems:~

nandhu@ai-systems:~$ ./profile_

Nandhu Raj R

Applied AI / LLM Systems Engineer

Building AI/software systems across model evaluation, RAG workflows, agentic AI orchestration, edge AI, data pipelines, and product-grade AI tooling.

AGENTIC AI MODEL EVAL EDGE AI AI SYSTEMS HEALTHTECH
[View Projects] [Resume] [GitHub] [LinkedIn]

$ whoami_

I recently completed my B.Sc. in Management & Technology at the Technical University of Munich, specializing in Computer Engineering.

My work sits at the intersection of AI systems, model evaluation, backend/data workflows, and product execution. Over the past 3+ years at Brainlab, I worked on R&D platforms, data and analytics pipelines, integrations, internal AI tooling, and production-facing prototypes.

My thesis focused on activation steering and behavioral evaluation in masked diffusion language models, including refusal, sycophancy, hallucination, transfer behavior, layer depth, and denoising-time effects.

$ system_status_

Brainlab R&D Systems

platforms, integrations, analytics, internal AI tooling

LLaDA Model Evaluation

activation steering and behavioral analysis workflows

Agentic AI Orchestration

agent workflows, tool routing, and production execution pipelines

Applied AI Products

from prototype to product-grade implementation

$ featured_systems_

Jarvis - Local-First Multimodal AI Desktop Assistant

BUILT / IN PROGRESS

Local-first desktop AI assistant for local and cloud LLMs with screenshot/image context, streaming chat, provider switching, secure key storage, voice/camera interaction design, and privacy-first controls.

Tauri Rust Svelte Ollama WebRTC SQLite Local AI Multimodal
repo

Activation Steering in LLaDA: Localization, Generalization, and Behavioral Structure

RESEARCH

TUM thesis on activation steering and behavioral evaluation in masked diffusion language models, analyzing refusal, sycophancy, hallucination, transfer behavior, layer depth, and denoising-time effects. Reduced refusal from 68.8% to 3.1% on the main harmful-prompt benchmark and localized strongest steering effects near layer 19 and the first 40 denoising steps.

Model Evaluation LLaDA Masked Diffusion LMs LLM Safety Behavioral Analysis
repo

StudyBridge - Agentic RAG Study Workspace

BUILT

Agentic RAG workspace for course-material ingestion, source-grounded tutor responses, summaries, flashcards, quizzes, and structured study artifacts. Implemented PDF/DOCX/PPTX/text ingestion, chunking, passage ranking, source-ID filtering, and multi-backend inference across local and cloud models.

RAG LangChain Ollama OpenAI Anthropic Gemini Source-Grounded QA
repo

Upgrant - Local AI Workflow Platform

BUILT

Local-first AI productivity system with multi-provider model routing, Codex CLI/Ollama/OpenAI integration, structured outputs, runtime checks, and human approval gates for AI write actions.

Electron React Ollama Codex CLI Structured Outputs Human-in-the-loop
repo

Crypt - AI Smart Contract Security Platform

BUILT

AI-assisted smart-contract security platform combining static analysis, exploit validation, LLM triage, structured outputs, monitoring, and audit-report generation. Built with FastAPI orchestration, Docker Compose services, Slither/Mythril/Foundry-style analysis workflows, multi-provider LLM analysis, and regression testing.

FastAPI Docker LLM Orchestration Security Foundry Monitoring
repo

$ research_experiments_

XAI Fraud Detection Pipeline

Reproducible XAI pipeline over SEC EDGAR disclosures using FinBERT, cross-validation, holdout evaluation, SHAP, LIME, attention, permutation, and counterfactual explanations. Achieved 80.0% accuracy, 93.33% recall, and 82.35% F1 on a 30-sample holdout set while preserving interpretability through explanation reports.

FinBERT XAI SHAP LIME Model Evaluation Financial NLP

Denoising-Step Efficiency in Masked Diffusion LMs

Research experiment studying how behavioral steering effects and output quality vary across denoising steps in masked diffusion language models. The LLaDA experiments showed that strongest refusal-steering effects were concentrated near layer 19 and the first 40 denoising steps, suggesting that useful behavioral control may not require uniform intervention across the full sampling trajectory.

Masked Diffusion LMs LLaDA Denoising Steps Inference Efficiency Model Evaluation

OpenAI Parameter Golf Challenge - Experimental Submission

Experimental small-model optimization submission to OpenAI's parameter-golf challenge, focused on compression, quantization, and evaluation under strict artifact-size constraints. Improved FineWeb validation BPB from 1.2244 to approximately 1.147 in a 15.8MB quantized artifact using quantization, EMA, int6 export, late QAT, partial RoPE, and embedding/activation changes.

Model Compression Quantization Evaluation FineWeb BPB Small Models

$ experience_log_

  • [2023-present] Brainlab - Data, AI & Software Systems, R&D Office

    Built R&D platforms, data pipelines, analytics workflows, integrations, and internal AI prototypes in a complex healthtech R&D environment.

  • [2026] TUM Thesis - Activation Steering in LLaDA

    Studied behavioral control in masked diffusion language models through controlled model-evaluation experiments.

  • [2025-2026] Crypt - Founder / Builder

    Built an AI-assisted security platform with backend orchestration, exploit-validation workflows, structured LLM outputs, monitoring, and audit-report generation.

$ product_and_strategy_

  • FEATRme - Associate Product Manager

    Worked on product discovery, AI-LegalTech feature planning, customer interviews, prototyping, GDPR considerations, and funding-pitch preparation.

  • SeniorConnect - Founder's Associate / GTM Strategy

    Conducted B2B market research across SaaS, IT, and industrial sectors, supporting outreach, vertical analysis, and GTM strategy.

  • WearCompare - Co-founder / Product

    Built and validated an AI-powered clothing resale/reuse concept in collaboration with Circular Republic, conducting 50+ qualitative interviews and launch strategy work.

  • PushQuantum - Corporate Sponsorship Coordinator

    Managed sponsorship and partnership development, including work connected to a BMBF-funded quantum outreach initiative.

$ technical_stack_

languages -> Python, TypeScript, JavaScript, SQL, Java, R

ai_ml -> PyTorch, TensorFlow/Keras, scikit-learn, Hugging Face, FinBERT, SHAP, LIME, model evaluation, RAG

llm_systems -> OpenAI API, Anthropic API, Gemini, Ollama, llama.cpp, Codex CLI, LangChain, LlamaIndex, Pinecone

backend -> FastAPI, REST APIs, WebSockets, Pydantic, Docker, GitHub Actions, CI/CD

frontend -> React, Next.js, Tailwind CSS, Vite, Zustand

data_cloud -> Power BI, Tableau, Azure DevOps, Azure, AWS

$ leadership_

  • IIT Madras Leadership & Innovation Roles

    Held student leadership roles across academic affairs, cultural affairs, student innovation, entrepreneurship, department representation, and campus media.

  • Academic Affairs · CFI · E-Cell · Department Council · Cultural Affairs · Article 19

    Academic Affairs Executive Committee, Centre for Innovation Coordinator, E-Cell Associate Manager, Department Council Member, Cultural Affairs Executive Council, and campus media.

  • UN / Public Health Volunteering

    Supported BRUNCH and WASH outreach initiatives targeting child nutrition and public-health awareness, contributing to digital outreach and campaign growth.

$ education_

  • [2022-2026] Technical University of Munich

    B.Sc. Management & Technology - Specialization: Computer Engineering

  • [2020-2022] Indian Institute of Technology Madras

    Integrated M.A. Interdisciplinary Studies - coursework completed before transfer to TUM

$ resume_

resume_status: unavailable

Resume is currently not published here. Please connect via LinkedIn or GitHub for the latest version.

$ connect_

nandhu@ai-systems:~$ connect

GitHub: github.com/NandhuRajRK LinkedIn: linkedin.com/in/nandhu-raj

Location: Germany