SynaptiQ: Building a Local-First Conversational Analytics Platform with NL-to-SQL, Forecasting, and Offline Voice Support

Abstract

Here’s what happened. Thirteen steps shaped this thing called SynaptiQ, built during our last year as engineers. It started because we got tired of seeing folks stuck with data in spreadsheets or CSVs - data they could not explore unless they learned SQL or handed everything over to an online service. That felt wrong. Our version works only on your machine. Nothing leaves your device. Questions in everyday language go in. A model on your own device turns them into code that talks to a built-in database. The result comes back with a visual graph attached. Data gets checked automatically when it arrives. Predictions over time happen through methods like SARIMAX along with alternatives. You can test changes to see possible outcomes. Voice works both ways if you choose it. All of it stays inside the computer. At point M12, forty-eight checks were cleared. Real data worked smoothly. No link to the web was needed. Labels used: talking to databases with words, offline reasoning engines, compact analytical storage, keeping insights private, trend guessing math, exploring change effects, safe query rules, chat-style number crunching, self-hosted intelligence---runners, memory-backed-search-tools.

Introduction

The text describes a locally run, privacy-focused data analytics system designed for non-technical users who need to extract insights from data without relying on external servers or coding expertise.

The core problem is that existing tools require SQL knowledge or cloud uploads, which creates usability barriers and serious privacy risks, especially in sensitive domains like healthcare, law, and finance. To address this, the system is built to run entirely offline with full transparency, where every SQL query and decision process is visible and explainable.

The architecture is modular, consisting of layers such as a core engine, data processing layer, analytics modules (querying, forecasting, visualization, insights, and what-if analysis), and an API interface connected to a frontend. A strict pipeline ensures all communication flows through controlled layers, with guardrails enforcing safe SQL execution before queries reach DuckDB.

The system supports multiple capabilities: natural language to SQL conversion using LLMs with validation safeguards, file upload with hashing to avoid duplication, automated data cleaning, hybrid storage using DuckDB and vector databases (ChromaDB), forecasting using statistical models like SARIMAX, and scenario-based what-if analysis. It also includes voice interaction powered by Whisper-based speech recognition and text-to-speech output.

Conclusion

We built SynaptiQ to solve a real problem we observed: people with useful data and no way to ask questions about it. The constraint we imposed on ourselves — everything runs locally, nothing leaves the machine — turned out to shape the entire architecture in interesting ways. Over thirteen milestones we got from a rough prototype to something that handles real datasets, produces useful forecasts, lets users model scenarios, and does all of it without an internet connection. Every test in the 48-part series ran without issue. Safety checks inside SQL stayed strong under pressure. Boundaries around modules? They lock into place by themselves. Testing proved it works. Even now, work remains. Think retry loops, embedding in batches, logging in securely. Yet what’s built already stands firm. The heart of it works. Data stays close. Yours alone. Clear to see how it runs. Never sent off to some distant machine.

References

[1] B. Qin et al., “A Survey on Text-to-SQL Parsing: Concepts, Methods, and Future Directions,” arXiv:2208.13629, 2022. [2] V. Zhong, C. Xiong, and R. Socher, “Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning,” arXiv:1709.00103, 2017. [3] X. Xu, C. Liu, and D. Song, “SQLNet: Generating Structured Queries Without Reinforcement Learning,” arXiv:1711.04436, 2017. [4] T. Yu et al., “Spider: A Large-Scale Human-Labeled Dataset for Text-to-SQL,” EMNLP 2018. [5] B. Wang et al., “RAT-SQL: Relation-Aware Schema Encoding for Text-to-SQL,” ACL 2020. [6] T. Scholak, N. Schucher, and D. Bahdanau, “PICARD: Parsing Incrementally for Constrained Decoding,” EMNLP 2021. [7] A. Radford et al., “Robust Speech Recognition via Large-Scale Weak Supervision,” ICML 2023. [8] Statsmodels Developers, “SARIMAX,” Statsmodels Documentation, 2024. Available: https://www.statsmodels.org [9] M. Raasveldt and H. Muhleisen, “DuckDB: An Embeddable Analytical Database,” ACM SIGMOD, pp.1981–1984, 2019. [10] M. Kleppmann et al., “Local-First Software: You Own Your Data,” ACM SPLASH ONWARD!, 2019.

Copyright

Copyright © 2026 Swetha P M, Likhith C P, Tanish S, Ganga L, Suman M N . This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.