I design and build practical AI systems — from machine learning experiments and RAG pipelines to data-driven products that turn messy information into usable decisions.
About
I'm Sanjay Ram A, an AI engineering student from Tamil Nadu building real products on the side. I started with curiosity about how models actually work — and ended up shipping apps, fine-tuning LLMs, and founding Dotwellabs.
My work sits at the intersection of applied machine learning, mobile AI, and product thinking. I care about systems that are measurable, reproducible, and useful outside a notebook — not just impressive in a demo.
When I'm not training models or debugging inference pipelines, I'm writing about what I learn, contributing to open source, and looking for problems worth solving with AI.
Models and datasets published on Hugging
Face
Selected Work
AnyLLM is the ultimate AI chat app — one unified workspace to talk to GPT-4o, Claude, Gemini, Llama, Groq, and more. Stop juggling five different apps. Bring your own API key and get the most out of every top AI model in one place. Supports Web search, MCPs, AI-Debates, Custom providers via OpenAI-compatible APIs and Agentic skills.
AI-powered appliance repair assistant, built entirely on-device for the Gemma 4 Good Hackathon. FixGemma is a Flutter application that leverages Gemma 4 (e2b & e4b models) via the Cactus AI inference engine to provide users with private, completely offline, multimodal repair assistance.
An interactive platform that enables AI agents from different domains to engage in structured debates on various topics. it is a Flask-based web application that facilitates debates between AI agents specialized in different domains. Each agent utilizes domain-specific knowledge bases and Large Language Models (LLMs) to generate informed responses and engage in meaningful discussions.
Offline mobile TTS pipeline built around Kyutai’s Pocket TTS model via ONNX runtime (sherpa-onnx), with zero-shot voice cloning fully on-device. Record a 5–12s Voice Profile or use built-in voices, tune inference parameters for speed vs quality — fast synthesis with no network dependency and complete privacy.
Pretrained a 15M-parameter Llama-style decoder-only transformer from scratch on a single RTX 3050 (4GB VRAM) across ∼1B tokens, with a custom 6K byte-level BPE tokenizer and staged SFT for chat and tool-calling. Final validation perplexity ∼53.4 at under 1.1GB peak memory; open-sourced model and full pipeline.
Technical Stack
01
02
03
04
Interests
Finetuning, RAG, agents, evaluation, prompt design, and reliable AI workflows.
Dataset curation, supervised fine-tuning (SFT), parameter-efficient tuning, and rigorous evaluation.
Dashboards, APIs, and interfaces that make model outputs understandable.
Metrics, error analysis, reproducibility, and honest measurement.
Blog / Notes
Kaggle Gemma hackathon writeup — fine-tuning E4B/E2B, running fully offline on a budget phone, and shipping a carousel UI.
I’m interested in AI engineering roles, ML projects, research collaborations, and product ideas where machine learning needs to be useful, measured, and shipped.