System
Models live where your data lives.
A confidential meeting assistant with speaker-aware transcription, strategic analysis, and competitive intelligence. Zero data leakage, fully on-premise.
I
Confidentiality as a non-negotiable constraint
Models live where your data lives. No external APIs, no uploads, no shared cloud. Atena runs on-premise on your servers, and every architectural decision is subordinate to this principle. The critical path never calls a service outside your perimeter.
II
Speaker-aware transcription and diarization
Multi-channel audio capture, deep-learning noise suppression, Italian transcription above 95% accuracy on technical jargon, speaker recognition via prior voice enrollment. The meeting ends — the transcript with the right names is ready within minutes.
III
Strategic analysis and SWOT
Three sequential passes of an on-premise LLM grounded in the context documents: factual summary, SWOT anchored to the technical material, mapping of criticalities and gaps. No free hallucination: every claim cites its source.
IV
Privacy-preserving web intelligence
Outbound queries pass through three filters: rule-based sanitizer, LLM rewrite into generic terms, mandatory human approval gate. Execution on a self-hosted SearXNG. For the most cautious customers, a fully offline mode on a pre-indexed corpus.
V
Signed report with audit trail
Final output in DOCX and PDF, customer-branded, with executive summary, transcript, SWOT, competitive intelligence with sources, action items and an immutable decision log. Digital signature with the customer's own certificate.
Capabilities
01
Multi-channel audio capture
Tabletop microphone array or individual lavaliers. WAV 48 kHz / 16 bit, one file per channel.
02
Deep-learning noise suppression
DeepFilterNet3 in real-time. Removes ambient noise, fillers, keyboard clicks.
03
Italian ASR above 95% accuracy
Whisper large-v3 via faster-whisper, optional fine-tuning on the customer's lexicon.
04
Speaker identification
ECAPA-TDNN voice prints in pgvector. Active learning after the meeting.
05
RAG on context documents
Multilingual embeddings, hybrid dense + BM25 search, dedicated reranker.
06
GDPR and explicit consent
Per-participant consent workflow. Configurable retention, segment exclusion.
07
End-to-end encryption
LUKS at-rest, mTLS in-transit. Keycloak + RBAC for granular access to meetings.
08
Immutable audit log
Append-only on Postgres, hash chain. Every decision, query, export traced.
Technology stack
All open-source or permissively licensed. No vendor lock-in. No closed models. No cloud service in the critical path.
- WhisperXASR + forced alignment
- pyannote.audioSpeaker diarization
- DeepFilterNet3Noise suppression
- ECAPA-TDNNVoice fingerprint
- BGE-M3Multilingual embeddings
- pgvector + PostgresVector store + DB
- vLLM + Qwen3 / Llama 3.3On-premise LLM runtime
- SearXNGSelf-hosted meta-search
- KeycloakEnterprise SSO
- Docker / K3sOn-premise deploy
Delivery roadmap
Four phases with concrete deliverables. Each phase is self-contained: the customer sees value before approving the next.
01 · 2 weeks
Discovery
IT environment audit, sensitive entity lists, technical corpus collection, operational requirements finalization. Signed Statement of Work.
02 · 4–5 weeks
MVP Pipeline
End-to-end pipeline: audio → denoise → transcription → anonymous diarization → synthesis → markdown report. Working demo on a test meeting.
03 · 3 weeks
Speaker ID + RAG
Voice enrollment app, voice prints database, runtime identification, RAG over context documents, document-aware SWOT pass.
04 · 6 weeks
Web Intelligence + Hardening
Sanitizer + approval gate, SearXNG, results integration, penetration test, operator training, go-live.