Built at IIFL Samasta

PresntAI

Enterprise Meeting Automation Platform (On-Premises)

Audio AI | Speech Processing | Enterprise Meeting Automation

9 Technologies·7 Key Features·2026·high complexity

Overview

PresntAI is an on-premises, end-to-end meeting intelligence platform that converts raw meeting recordings into speaker-attributed, context-aware documentation automatically. It handles the full pipeline: audio standardization, transcription, speaker diarization, speaker identification, transcript alignment, and AI-driven meeting analysis. The system generates Minutes of Meeting (MoM), speaker-wise action points, executive summaries, and maintains an understanding of previous meetings for continuity. Entirely hosted on internal infrastructure, ensuring enterprise privacy.

Key Features

On-Premises & Privacy-Preserving

All processing, storage, and distribution run inside the corporate network. No external SaaS.

Complete Audio-to-Documentation Pipeline

From raw audio (up to 5 GB) to polished PDF reports: transcript, diarization, MoM, decisions, commitments.

Speaker Diarization and Real-Name Identification

pyannote diarization + SpeechBrain ECAPA embeddings with enrolled speaker database.

Speaker-Wise Action Points & Contextual Memory

AI extracts decisions and action items by speaker, leveraging previous meeting context.

Automatic MoM Generation & Email Distribution

Professional MoM automatically generated and emailed to all participants.

In-UI Recording

Built-in recording eliminates dependency on external tools.

Multi-LLM Analysis

Vicuna:13b (noise cleanup), DeepSeek R1:14b (outcomes), Llama3 (summary).

Technology Stack

Algorithms & Methods

Audio Standardization

FFmpeg: input → 16kHz mono WAV.

ASR (Speech-to-Text)

faster-whisper with VAD filter, beam size 10.

Speaker Diarization

pyannote pipeline generating labeled speaker segments and RTTM.

Speaker Attribution

Transcript chunks mapped to diarization segments by timestamp overlap.

Speaker Recognition

ECAPA embedding → cosine similarity against enrolled DB → threshold-based Unknown fallback.

Multi-LLM Intelligence

Noise cleanup (Vicuna), Outcomes (DeepSeek R1), Summary (Llama3), Speaker Analysis (DeepSeek R1).

System Architecture

API Server

Upload, analysis, status, results, report serving, health endpoints

Orchestration Engine

Full pipeline: transcription → diarization → alignment → LLM analysis

Diarization Module

pyannote diarization, RTTM + readable log output

Speaker Transcript Builder

Transcript + diarization merge, speaker-attributed output

Speaker Recognizer

Enrolled speaker DB, ECAPA inference, cosine similarity matching

Meeting Analysis Pipeline

Ollama prompt execution, markdown sanitization, PDF generation

Results & Outcomes

  • Full on-premises meeting intelligence platform operational at IIFL Samasta
  • Speaker-attributed transcripts with real name mapping
  • Automatic speaker-wise action points and contextual memory across meetings
  • Zero manual documentation and zero manual distribution
  • Windows-stable deployment using faster-whisper + pyannote + Ollama

Related Projects