Team03 -SLAM : Small Language Agentic Machine

Indian Institute of Science Summer 2025 DA225o Submission21 Authors

Keywords: Small Language Models, Tool-Augmented AI, Task Classification, LoRA (Low-Rank Adaptation), Modular AI Systems, Edge AI, Agentic AI Systems, Prompt Engineering, Decoder-Only Models, Encoder-Decoder Models, Lightweight Orchestration, Structured Prompt Generation

TL;DR: SLAM is a lightweight, local-first AI assistant that interprets user input, chooses tools, and completes tasks without relying on cloud services or heavy frameworks

Abstract: SLAM (Small Language Agentic Machine) is a lightweight, local-first assistant that is designed to perform practical tasks through modular tool execution and coordinated language-model reasoning. SLAM adopts a two-stage architecture: an encoder-decoder T5 model first rewrites user inputs into concise, instruction-like prompts; a decoder-only SLM then interprets these prompts, decides whether a tool call is needed, and generates the final reply. A lean in-process controller validates tool JSON, executes a sandboxed registry (like calculator, OCR, summarizer, formatter, …), and injects each result back into the generator. The entire stack is self-contained, requires no cloud APIs, and can be adapted to new domains via adapter-based fine-tuning of the rewriting model. By decoupling prompt interpretation, dialogue generation, and task execution, SLAM achieves agent-like behavior while remaining efficient, extensible, and deployable on lightweight hardware.

Submission Number: 21