# BiltIQ AI

> BiltIQ AI (a brand of Aarna Tech Consultants Pvt. Ltd.) architects, deploys,
> and maintains on-premise agentic AI systems for Indian enterprises in
> Healthcare, BFSI, Education, Manufacturing, and Government. Zero cloud
> dependency. 100% data sovereignty. DPDP-compliant.

Updated: 2026-08-24. Full page content for LLM ingestion: https://www.biltiq.ai/llms-full.txt

## Core

- [Home](https://www.biltiq.ai/): What we do and our value proposition.
- [About](https://www.biltiq.ai/about): Team, mission, parent company Aarna Tech Consultants Pvt. Ltd. Founded 8 September 2021 by Harish Kumar and Sonali Harish in Jamshedpur, Jharkhand.
- [AI Factory](https://www.biltiq.ai/ai-factory): Private, full-stack enterprise AI on office-grade hardware — agents over your own CRM, documents, and support data, with nothing leaving the building.
- [DPDP Compliance](https://www.biltiq.ai/dpdp-compliance): How India's DPDP Act 2023 applies to enterprise AI — and why on-premise deployment satisfies it by construction.
- [Contact](https://www.biltiq.ai/contact): Book a 30-minute architecture consultation. Email hello@biltiq.ai. Phone +91 94313 05522.
- [Careers](https://www.biltiq.ai/careers): Open engineering roles — AI/ML, Full-Stack, Voice/NLP, DevOps.

## Services

- [Agentic AI](https://www.biltiq.ai/services/agentic-ai): On-premise multi-agent AI with progressive autonomy — agents that plan, use tools via MCP, and act on your data with full audit trails.
- [Custom AI-Native Apps](https://www.biltiq.ai/services/custom-ai-apps): Custom AI-native applications on your infrastructure — on-premise LLM deployment, RAG pipelines, MCP integration, and fine-tuning.
- [Agents-as-a-Service](https://www.biltiq.ai/services/agents-as-a-service): Subscribe to AI agents, not SaaS licences — personal, team, and enterprise agent fleets on your infrastructure. Fixed cost, unlimited usage, zero per-query fees.
- [AI Orchestration](https://www.biltiq.ai/services/ai-orchestration): On-premise AI orchestration — GPU provisioning, vLLM multi-model serving, MCP routing, and Qdrant vector search.
- [Document & Conversational AI](https://www.biltiq.ai/services/document-ai): RAG pipelines, knowledge bases, and multilingual agents that cite answers from your own data.
- [Voice & Video Intelligence](https://www.biltiq.ai/services/voice-intelligence): Voice agents, ASR/TTS pipelines, and video intelligence entirely on your infrastructure — sub-200 ms latency, 100+ languages.
- [AI Factory Advisory](https://www.biltiq.ai/services/ai-factory-advisory): Helps Indian CIOs, CTOs, and boards decide on-premise vs hybrid vs cloud AI architecture — model, GPU sizing, and hosting — before committing hardware or contracts.
- [Structured Data for AI](https://www.biltiq.ai/services/structured-data-for-ai): Agent-ready data engineering — on-premise ingestion, ontology design, MCP-compatible tool contracts, continuous-sync governance.

## Products

- [ATC Manthan](https://www.biltiq.ai/products/atc-manthan): On-premise document intelligence agent. PDFs and scans → cited Q&A, study guides, audio overviews. 100+ languages, zero hallucination.
- [ATC Quest LMS](https://www.biltiq.ai/products/atc-quest-lms): Enterprise on-premise LMS with five AI learning agents. SCORM/xAPI/LTI; 1M+ concurrent users; works offline.
- [BiltIQ Campus](https://www.biltiq.ai/products/ai-campus): NEP 2020-aligned K-12 AI platform. 204 curriculum modules across 4 age groups. Works without internet.
- [ATC Voice](https://www.biltiq.ai/products/atc-voice): On-premise voice agents with sub-200 ms latency and 100+ language support.
- [ATC Chat](https://www.biltiq.ai/products/ai-chatbots): RAG + MCP conversational AI agent across web, WhatsApp, Telegram, and voice channels.
- [ATC Flow](https://www.biltiq.ai/products/atc-flow): MCP-powered data pipelines and workflow orchestration on local LLMs.
- [ATC CMS](https://www.biltiq.ai/products/atc-cms): Seven autonomous content agents — writer, editor, SEO, publisher, analyst, scheduler, optimiser.
- [ATC Connect](https://www.biltiq.ai/products/atc-connect): Process intelligence agent with 500+ integrations and AI decision-making.
- [ATC Social](https://www.biltiq.ai/products/atc-social): AI-powered social media management with content, scheduling, and analytics agents.
- [ATC Ops](https://www.biltiq.ai/products/devops-tools): AI infrastructure operations — GPU monitoring, predictive maintenance, auto-scaling.

## Industries

- [Healthcare](https://www.biltiq.ai/industries/healthcare): DPDP- and HIPAA-aligned clinical AI for hospitals and labs.
- [Banking & Finance / BFSI](https://www.biltiq.ai/industries/finance): On-premise AI for KYC, fraud triage, compliance — DPDP, RBI, ISO 27001.
- [Education](https://www.biltiq.ai/industries/education): Adaptive tutoring and NEP 2020-aligned LMS for K-12, universities, and corporate L&D.
- [Manufacturing](https://www.biltiq.ai/industries/manufacturing): SOP search, defect logs, predictive maintenance for shop-floor AI.
- [Government](https://www.biltiq.ai/industries/government): Air-gapped sovereign AI for citizen services and policy intelligence in 22+ Indian languages.
- [SMBs](https://www.biltiq.ai/industries/smbs): Pre-built agent bundles deployable in 3–5 days at 60–80% lower cost than SaaS.

## Content

- [Blog](https://www.biltiq.ai/blog): Engineering deep-dives on on-premise agentic AI, LLM fine-tuning, RAG, MCP, and DPDP-compliant deployment.
- [Blog RSS feed](https://www.biltiq.ai/blog/feed.xml): Machine-readable feed of the latest articles.
- [White Papers](https://www.biltiq.ai/whitepaper): Free white papers on on-premise AI architecture, DPDP compliance, and on-prem vs cloud TCO.
- [Media Room](https://www.biltiq.ai/media): Featured stories, news, awards, speaking engagements, and press kit.
- [India AI Summit 2026](https://www.biltiq.ai/india-ai-summit-2026): Visit BiltIQ AI at Booth 135, Hall 1, Bharat Mandapam, New Delhi (Feb 16–20).

## Resources

- [When you should use Claude or GPT instead](https://www.biltiq.ai/blog/when-to-use-claude-or-gpt-instead): We build on-premise AI infrastructure and this sets out when buying it would be a mistake. Below roughly 90,000 queries a month no on-premise deployment pays back inside three years, including our entry tier. Covers the five cases where cloud frontier models are the correct purchase, four things that do not disqualify you, and the hybrid split.
- [Shadow AI: your staff are already pasting your IP into chatbots](https://www.biltiq.ai/blog/shadow-ai-staff-pasting-company-data): Two-thirds of surveyed office professionals used AI at work believing it was not permitted under company policy, and 43% entered work correspondence into public tools. Why policy does not fix it, why smaller organisations are more exposed than larger ones, and what actually removes the incentive. Survey covers AU/JP/UK/US, not India, and says so.
- [Multi-agent architecture for unstructured business data](https://www.biltiq.ai/whitepaper/multi-agent-architecture): The engineering companion to the retrieval-first argument. Each layer of a business-data AI system explained as a response to a way business data breaks: layout-aware parsing, hybrid BM25F and dense retrieval fused by reciprocal rank fusion with a multimodal reranker, five-layer memory including a knowledge graph and audit trail, role-scoped agents under step and token budgets with Ed25519-signed delegation, and injection control. Each section states what that layer costs and what it still fails at, including that aggregation questions are not retrieval questions and that no injection scanner is complete.
- [Operational alpha: the advantage you own](https://www.biltiq.ai/whitepaper/operational-alpha): Why the corrections, vocabulary and institutional judgement an organisation feeds its AI system compound into an owned asset on-premise, and leak or expire in a rented one. Includes the volume economics: below roughly 90,000 queries a month no on-premise deployment pays back inside three years; a Rs 41.7 lakh reference fleet crosses over in about 18 months at 1 million queries a month with 700-token contexts.
- [Frontier vs grounded: the wrong axis](https://www.biltiq.ai/whitepaper/frontier-vs-grounded): Why model capability is the wrong axis for grounded enterprise work — output quality is dominated by what the system read, not model size — and the honest list of cases where frontier models win: no private grounding, low volume, spiky workloads, frontier-only capability. Sovereign by default, frontier by choice: commercial APIs policy-gated, default-off, logged.
- [Three ways to deploy enterprise AI](https://www.biltiq.ai/whitepaper/deployment-approaches): Cloud API, hybrid and on-premise compared with every figure carrying its volume — the 90,000-queries-a-month floor below which cloud wins, crossover at 8–14 months from about 0.4M queries a month, and running costs measured on a real fleet (about Rs 35,000 a year typical draw against about Rs 8.9 lakh for one 8xH100 cloud-class server).
- [The anatomy of an on-premise AI deployment](https://www.biltiq.ai/whitepaper/anatomy-of-an-on-premise-deployment): What an on-premise AI deployment physically is and how it operates. A departmental deployment is one desk-side unit drawing about 240 W, because inference is not training. Covers the three hardware tiers with capex, the software that runs on them, how documents are ingested with their existing access permissions mirrored into the index, the request path, who operates it, what breaks, and what a deployment does not include. The hardest part is inheriting the permission model, not the hardware or the model.
- [Your documents are worse than you think](https://www.biltiq.ai/blog/your-documents-are-worse-than-you-think): A document that failed to parse is indistinguishable from a document containing nothing relevant, and no error appears. Six ways real business documents defeat naive extraction — scans without a text layer, flattened tables, merged cells, quoted email history, nested containers, diagrams that are pictures — what layout-aware parsing does about it, what still does not work, and the five-worst-documents test to run before buying anything. Publishes no parsing accuracy percentage and says why.
- [Why RAG beats fine-tuning for business knowledge](https://www.biltiq.ai/blog/why-rag-beats-fine-tuning-for-business-knowledge): Fine-tuning teaches a model how to behave; retrieval supplies what is true. Business facts change, need citation, need access control and need deletion — four properties fine-tuned weights cannot provide. States fine-tuning's real uses fairly: output format, domain vocabulary, task shape, latency at scale. The recommended design is both together: behaviour in the weights, facts in the index.
- [What your DPO will ask about your AI vendor](https://www.biltiq.ai/blog/what-your-dpo-will-ask-about-your-ai-vendor): Ten procurement questions written to be pasted into a vendor pack and used verbatim, including against BiltIQ. The distinction that runs through all ten: a control you can describe is a policy, a control that emits a record every time it fires is evidence. Makes no regulatory claim — counsel owns that question.
- [One agent can't do it: why orchestration matters](https://www.biltiq.ai/blog/why-one-agent-cannot-do-it): The most common failure in agentic deployments is one general agent with every tool and no budget. It wanders, loops, over-reaches and cannot be debugged — each a symptom of a specific missing constraint: step budgets, circuit breakers, scope isolation enforced below the agent, named roles, authenticated delegation. Also states when one agent is right and what orchestration costs.
- [The compensating-control trap](https://www.biltiq.ai/blog/the-compensating-control-trap): "We encrypt it and sign a DPA" is a real control but not the same thing as processing that never leaves the building. Compensating controls recur on someone else's cadence, are exercised at arm's length, and produce assurance rather than evidence. Describes how controls behave, not what any regulation requires, and states where the argument does not apply.
- [The five-layer memory model](https://www.biltiq.ai/blog/the-five-layer-memory-model): Working, session, project, knowledge graph and audit trail — each layer has a different retention period and a different access-control boundary, and systems that conflate them leak. Covers promotion between layers as an explicit decision, entity resolution as the unsolved hard part, and the question to ask of any system offering memory: which layer is this remembered in, and who else can read it?
- [What actually runs on the box?](https://www.biltiq.ai/blog/what-actually-runs-on-the-box): An on-premise AI node runs six things, and only one of them is a model: a vLLM inference server, a model registry mapping roles to pinned versions, vector and lexical indexes, an ingestion pipeline that captures document entitlements at ingestion, an agent runtime with budgets, and the controls (privacy filter, injection scanner, output validator, hash-chained audit trail). Application logic usually lives off the box. The model is the replaceable part; the index, entitlement mapping and audit chain carry forward.
- [Why inference-time engineering is the whole game](https://www.biltiq.ai/blog/why-inference-time-engineering-is-the-whole-game): The reason a modest node serves a department at interactive speed is inference-time engineering, not the accelerator. Continuous batching and paged attention (the platform's serving layer) explained plainly, plus the field's wider toolkit — quantisation, KV-cache compression, speculative decoding, prefix caching, expert offloading — described as the state of the art, not claimed as features. Publishes no throughput numbers and says why: serving performance is a property of model, quantisation, hardware and request mix together.

## Optional

- [Privacy Policy](https://www.biltiq.ai/privacy-policy)
- [Terms of Service](https://www.biltiq.ai/terms-of-service)
- [Refund and Cancellation Policy](https://www.biltiq.ai/refund-and-cancellation-policy)