Intelsense AI

.Technology.

The stack underneath every product

Speech, language and vision models built in house over eight years. Not licensed from a vendor, and not fine-tuned on top of someone else's base.

.Capabilities.

What the stack does

Grouped by the models they run on, rather than listed as a menu of services.

Agentic AI

Runs on Ekush LLM + SenseVoice + SenseSpeech

Agents that hold a conversation, make a decision inside a defined boundary, and carry out multi-step work, with a named human approving anything that leaves the building.

  • Lifecycle agents: find, convert, onboard, retain, support, collect
  • Shared context across voice, chat and web
  • Policy engine with default deny and named approvers
  • Tamper-evident audit chain on every action

Speech AI

Runs on SenseVoice + SenseSpeech

Recognition and synthesis for languages the global platforms treat as an afterthought, including the mixed Bangla-English that real customers actually speak.

  • Speech to text across 41+ languages
  • Text to speech with custom voice training
  • Speaker identification and noise handling
  • Voice-controlled navigation and transactions

Document AI

Runs on Vision models + Ekush LLM

Reading printed and handwritten Bangla and English with no template registered in advance, including the photocopy of a photocopy shot on a phone at an angle.

  • Sort by document type without a rule configured first
  • Split a scanned batch into individual records
  • Extract fields at 95–99% accuracy, printed and handwritten
  • Push structured data into core systems, no re-keying

Language AI

Runs on Ekush LLM

Understanding, generation and reasoning over text in languages where an English-first model loses the meaning.

  • Intent and sentiment on real customer language
  • Summarisation and extraction from long documents
  • Domain fine-tuning on your own data
  • Retrieval over your knowledge base

Why owning the models matters to you

A vendor reselling somebody else's API cannot put the intelligence inside your building, cannot promise your data stays in country, and cannot stop the model underneath them changing or being deprecated.

We can, because the models are ours. On premise in your data centre or private cloud you control, with every extraction traceable to its source field and every interaction logged. For a regulated buyer this is usually the whole decision.