Samarth Kadam
AI/ML engineer and researcher. Voice agents, swarm simulations and local-first AI tools.
About
I am a B.Tech student at RV University, Bengaluru, working on applied AI: conversational avatars that run fully offline, retrieval with vector embeddings, and digital twins for drone swarms.
I led RUDRA, the university’s data science and AI/ML club, and helped found its FOSS United chapter. I like building things people can actually use, and getting other people building too.
Education
B.Tech, with a minor in Innovation, Entrepreneurship & Product Development. CGPA 9.16 / 10 (fifth semester).
Research
H. J. Vrishank, Samarth Kadam, Harish K. R., Rammohan Bhanumurthy · RV University · To appear in proceedings from ADE (DRDO)
A digital twin that watches a UAV swarm as a graph and reconfigures it when things go wrong. Each drone only talks to neighbours in range; when connectivity drops below a threshold, the swarm pulls itself back together.
- UAV failure recovery (> 95 % mission success): Drones drop out mid-mission and the rest close the gap.
- GPS spoofing (≈ 100 % mission success): A spoofing zone corrupts positions; the twin spots the mismatch and corrects it.
- Connectivity restoration (≈ 88 % mission success): Jamming splits the swarm; the control law pulls the groups back into one.
- Threat avoidance (≈ 96 % mission success): A threat appears; an exclusion zone is drawn and the swarm replans around it.
How cities can design AI with the people who live with it, not just for them.
Co-authored chapter in an edited volume of current engineering research.
Projects
An empathetic talking avatar for officer wellness support that runs entirely offline.
- Built a multimodal pipeline for Bengaluru Police: speech synthesis (F5-TTS) plus video generation produce a realistic talking avatar.
- Runs LLaMA 3 locally through Ollama, so sensitive officer data never leaves the machine, with sub-second responses.
- Zero-shot voice cloning and lip-sync (SadTalker / Wav2Lip) turn the model’s text into matched audio and video.
- A Gradio interface manages the whole path from typed input to a rendered digital human on local hardware.
Lost-and-found by visual search: photograph an item and vector embeddings find its match.
- Retrieval system that uses high-dimensional vector embeddings to identify lost property from a photo.
- A real-time similarity scoring engine cuts manual recovery time.
- PostgreSQL backend for spatial and visual metadata.
- Android beta distributed through Firebase.
A lightweight desktop AI research assistant that runs local models through Ollama.
- GPU-friendly Tauri app with deep reasoning and tool-augmented answers, all on local inference.
- Live web search (Tavily / SearXNG) and multi-format file analysis for up-to-date research.
- Token-optimised "smart memory": custom chunking and persistent context keep long sessions efficient.
Six anomaly-detection models compared across twelve scientific-workflow datasets.
- Benchmarks 6 models on 12 FlowBench / cFlow-Bench datasets covering tabular, graph and text data.
- Supervised baselines (Random Forest, GCN, RoBERTa) and unsupervised ones (GMM, graph autoencoder, Sentence-BERT + Isolation Forest), all reported with Accuracy, F1 and ROC-AUC.
- Fully scripted: dataset download, training, evaluation, and per-model reports and plots.
20,000+ dermatoscopic images unified from ISIC and HAM10000 into one benign / malignant dataset.
- Aggregated and pre-processed 20,000+ images from ISIC and HAM10000 for binary classification.
- Cleaning and re-categorisation pipelines turn multi-class clinical labels into a balanced training resource.
- Standardised formats and metadata from six repositories to help models generalise across conditions and lighting.
- Published on Kaggle under CC BY-NC-SA 4.0.
Experience and leadership
- Technical team member from Nov 2024, President Jul 2025 – Jul 2026, Advisor since.
- Ran seminars and hands-on workshops for 500+ students.
- Launched a research platform that pairs students with faculty.
- Organised year-long hackathons and quizzes.
- Grew the campus open-source and blockchain community through mentoring.
- Ran 10+ workshops on decentralised technologies, reaching 1,000+ student developers.
- Organised coding sprints and guest lectures promoting free and open-source software.
- Co-founded the university’s FOSS United chapter.
- Maintains the university’s alumni-facing web presence.
- Connected product-development insights to marketing strategy across teams.
- Worked on brand positioning and team management during a fast summer internship.
Skills
- AI / ML
- PyTorch · PyTorch Geometric · Transformers · scikit-learn · LLMs · AI agents · RAG · Prompt engineering · Ollama · Speech synthesis (F5-TTS) · Vector embeddings · Computer vision
- Languages
- Python · TypeScript · JavaScript · Java · Rust · SQL · HTML / CSS · MATLAB
- Frameworks
- React · FastAPI · Tauri · Gradio
- Data
- PostgreSQL
- Tools
- Git · Hugging Face · Figma · UML design
Awards and certificates
- Winner, AI Build Sprint
- Runner-up, SAP Vibeathon
- Top 10 finalist, Tech Sprint Hackathon
- Quantum Computing
- Computer Networking Basics: How Computers Communicate
- Faculty Development Program (participation)
- Event Coordinator