Headline

Data Scientist at Astrikos AI

About

AI/GenAI Engineer with 2+ years building and shipping LLM-powered, agentic systems in production — from RAG assistants to autonomous action layers with real device control. I work end-to-end on GenAI systems: retrieval (naive top-k through hybrid dense + BM25 with cross-encoder reranking), tool-calling agents, multi-stage LLM orchestration, and — the part most people skip — rigorous evaluation of non-deterministic output. Some of what I've built: - A conversational text-to-SQL RAG assistant scoring ~91% SQL match on a 1,473-question eval suite (~1.9s avg response), with a self-correcting, multi-store FAISS retrieval layer. - An autonomous building-automation action layer over 1,000+ devices with gated, human-reviewed cloud actuation — writable-parameter allowlists, confidence-based shadow→active promotion, and rollback. - A hybrid-RAG document-QA engine (FAISS + BM25 + Reciprocal Rank Fusion + reranking) with cite-or-refuse grounded answers. - A LangGraph state-machine agent for document acquisition + structured extraction (Gemini Vision + Groq), containerized with Prometheus metrics. - An LLM distillation project fine-tuning Qwen 2.5 Coder 7B with QLoRA on BIRD-SQL. Toolkit: Python, LangChain, LangGraph, MCP, FAISS/Chroma/HNSW, Hugging Face, Groq, Google Gemini, FastAPI, PostgreSQL, Docker, plus classical ML (scikit-learn, XGBoost, Prophet).

Location

Bengaluru

AI stack

antigravity

claude code

gemini

hugging face

langchain

LlamaIndex

Work experience

Data Scientist at Astrikos AI, Apr 2026 - Present · 6 mos

Associate Data Scientist at Astrikos AI, Apr 2025 - Apr 2026 · 1 yr 1 mo, Bengaluru, Karnataka, India, - Ship production LLM/agentic systems for building operations: a text-to-SQL RAG assistant (~91% SQL match on a 1,473-question eval) and an autonomous action layer with gated, human-reviewed cloud actuation across 1,000+ devices. - Engineer self-correcting, multi-store FAISS retrieval and rigorous evaluation harnesses (sqlglot AST equality; offline temporal backtests) for non-deterministic LLM and control outputs. - Prototyped a Model Context Protocol (MCP) server exposing workspace tooling to LLM agents. - Built Remaining Useful Life (RUL) prediction and cascading failure-analysis pipelines for critical equipment (reported ~12.5% downtime reduction).

Data Science Intern at Astrikos AI, Jun 2024 - Apr 2025 · 11 mos, Bengaluru, Karnataka, India, - Built a no-code ML platform (drag-and-drop pipelines, automated feature engineering, HPO), cutting model-development time ~70% for non-technical users. - Added AutoML (regression/classification/clustering), vector-DB integration, and sensor failure-prediction models (~85% accuracy) with real-time streaming and explainable-AI reports.

Software Trainee at Hindustan Aeronautics Limited, Sep 2022 - Feb 2023 · 6 mos, Bengaluru, Karnataka, India

Intern at Exposys Data Labs, Aug 2022 - Sep 2022 · 2 mos

Campus Ambassador at Viral Fission, Apr 2021 - Mar 2022 · 1 yr

Student Partner at Internshala, Mar 2021 - Jan 2022 · 11 mos, India

Education

Master of Engineering - MEng, Big Data, Manipal School of Information Sciences

Bachelor of Engineering - BE, Computer Science, Acharya

Public links

github: https://github.com/preethamgouda