ARIA: a voice AI tutor built for hosted and frontier LLMs
Netcloud Consulting built and manages ARIA, an English and Hindi study assistant for UPSC aspirants. Students speak or type a question and get an answer grounded in their study notes, with the source shown and the reply read aloud. It runs on hosted open-source models today, and frontier models such as OpenAI or Gemini can plug into the same workflow when a job calls for them.
Qwen LLM
OpenAIGemini
- Whisper
- Kokoro TTS
Qdrant
Ollama
n8n
Percentages are calculated from before and after measurements on the live system.
ARIA in action
The challenge
UPSC aspirants need answers they can trust, tied to what they are actually studying. A general chatbot can sound confident while going off-syllabus, and it gives no way to tell whether an answer came from the student's notes or the model's memory.
The goal was a tutor students could talk to naturally, in English or Hindi, on an AI stack that stays under Netcloud Consulting's control, with the freedom to use frontier models where they add value.
What we built
A single workflow serves both chat and voice. Five self-hosted services do the AI work, connected through n8n and reachable only on a private network.
| Job | Component | Role in ARIA |
|---|---|---|
| Reasoning and answers | Qwen3.5 via Ollama | Hosted LLM that writes every answer from the retrieved notes |
| Frontier model option | OpenAI or Gemini | Can replace or back up the hosted LLM in the same workflow, while embeddings, search and speech stay hosted |
| Speech to text | faster-whisper | Transcribes spoken questions in English and Hindi |
| Text to speech | Kokoro-82M | Reads answers aloud and picks a Hindi voice for Hindi replies |
| Embeddings | Qwen3-Embedding via Ollama | Turns study notes and questions into searchable vectors |
| Vector search | Qdrant | Finds the most relevant passages for each question |
| Orchestration | n8n | Runs ingestion and the voice and chat workflow end to end |
Hosted and frontier models, one workflow
Hosted models keep data under your control and suit steady, high-volume work. Frontier models bring extra reasoning depth to harder questions, such as answers that go beyond the indexed notes. Because the model is a swappable part of the workflow, you can choose per use case, or run a frontier model with a hosted fallback, without rebuilding retrieval or speech.
How a question is answered
- Capture: the student types, or speaks and Whisper transcribes.
- Retrieve: ARIA searches the indexed study notes and keeps only relevant passages.
- Answer: the hosted LLM responds from those passages, using short-term memory for follow-ups.
- Label: each reply shows where it came from.
- Speak: the text appears first, then Kokoro reads it aloud with a lip-synced avatar.
Built for trust, not just conversation
Notes only by default
ARIA answers from the student's notes. If they do not cover a question, it says so and suggests a related topic.
Every answer is labelled
Source chips separate notes from general knowledge, and a verify prompt appears when facts may have changed.
Stays on topic
Questions outside UPSC preparation are politely declined and redirected to a study topic.
Private by design
AI services have no public ports, credentials stay on the server and requests are rate limited.
English and Hindi
Automatic language detection, or the student picks. Notes are searched in English to improve Hindi retrieval.
Works on any screen
Purpose-built desktop, tablet and phone layouts, with voice controls in easy reach.
The results
Moving retrieval ahead of the model and trimming each passage cut prompts by 44% and made answers 54% faster. Replying with text first and voice second means students start reading before the audio begins.
ARIA now gives students a grounded, spoken tutor in two languages, with every answer labelled by source and an AI stack under Netcloud Consulting's control that can add frontier models whenever they are the better fit. The same pattern applies to support, training and internal knowledge. See our voice AI agents, enterprise RAG systems and agentic AI development services, or browse more case studies.
ARIA FAQs
What is ARIA?
ARIA is a voice and chat study assistant for UPSC Civil Services preparation, built and managed by Netcloud Consulting. It runs on a hosted LLM with Whisper speech recognition, Kokoro text to speech and a Qdrant vector database, on an architecture that also supports frontier models such as OpenAI and Gemini.
Does ARIA send questions to an external AI API?
Not today. The language model, embeddings, vector search, speech recognition and text to speech all run on infrastructure Netcloud Consulting hosts. A frontier model such as OpenAI or Gemini can be connected when a job needs it, as the main model or as a fallback.
Which languages does ARIA support?
ARIA understands and answers in English and Hindi, by text or voice. It can detect the language automatically or use the language the student selects.
Can a business build a self-hosted voice agent like ARIA?
Yes. The same pattern of a self-hosted LLM, Whisper, text to speech and RAG suits teams that want voice AI over their own documents with control over where data is processed, and frontier models can be added where they improve results.
Want voice AI on hosted models, frontier models or both?
Tell us your use case and we will scope a voice or chat agent around your documents, languages and data rules, with the right mix of hosted and frontier models.
Talk to our team