Forensic intelligence over meetings.
Most meeting tools transcribe. Felarity finds contradictions, classifies confrontations, scores credibility, and signs the evidence so it survives outside the room.
Live meeting intelligence
The moment a session starts, the browser captures audio in five-second chunks and streams them to our gateway. Whisper large-v3 transcribes on dedicated GPUs and returns aligned text within roughly a second of the spoken word. As each utterance lands, the deep council compares it to everything said earlier in the session, scoring whether the new statement contradicts, qualifies, or commits the speaker to something they previously denied.
Credibility arcs update per speaker as the meeting unfolds. Hedge density rises, certainty shifts, a promise is made and then quietly walked back two minutes later — these are surfaced to the side panel in real time, not buried in a transcript you have to re-read at midnight.
- Contradiction alerts with cited prior utterance
- Per-speaker credibility arc (rising / falling / stable)
- Hedge and certainty drift
- Unkept commitments and forgotten promises
- Council-flagged escalations from the industry panel
- Side-panel intel cards as evidence accumulates
The post-session forensic pipeline
When the meeting ends, the eleven-stage pipeline runs against the full session — not the chunked live stream, but the reconstructed whole. Each stage emits a progress event over SSE so the workspace UI can show what is happening in the moment, and each stage's output becomes input to the next.
post_concat— Concatenate the chunked audio back into a continuous, gap-aware recording for whole-session analysis.Required because chunk boundaries break diarization and acoustic features.post_diarize— Run pyannote speaker diarization against the full session to bind utterances to speaker identities with higher confidence than the live pass.post_acoustic— Extract speech rate, pause distribution, and stress markers per speaker.Inputs into the credibility model and the confrontation classifier.post_confrontation— Classify confrontation and response pairs: was a hard question asked, what was the answer's structure, did the speaker deflect, pivot, or commit.post_samples— Extract a four-second voice sample per speaker for the report, so reviewers can verify who said what against the transcript.post_attribute— Bind the contradictions detected during the live phase to the speakers identified by the full-session diarization, correcting any drift from the live attribution.post_nli— Re-score contradiction pairs with DeBERTa-v3 natural language inference on CPU, producing entailment, neutral, and contradiction probabilities that survive scrutiny outside the room.post_topology— Build a NetworkX graph of statements, contradictions, and confrontations, then classify the meeting's pattern: hostile cross-examination, evasive consensus, productive disagreement, or rehearsed alignment.post_council— Deep council re-analysis with full speaker context. The industry panel reviews the now-attributed evidence and produces the final analyst commentary.post_attest— Build an eight-node SHA-256 Merkle attestation chain over the audio, transcript, diarization, contradictions, NLI scores, topology, council output, and metadata, then sign the root with Ed25519.post_complete— Render the final report: contradictions with citations, credibility arcs, confrontation map, council commentary, and the attestation block.
Cryptographic attestation
Every completed session ships with an eight-node Merkle tree whose leaves are the SHA-256 hashes of the audio, transcript, diarization, detected contradictions, NLI re-scoring, topology graph, council output, and session metadata. The root is signed with Ed25519 using our published signing key. The signature, the public key fingerprint, and the leaf hashes travel inside the report itself.
Any third party — opposing counsel, a regulator, an auditor, an arbitrator — can take the report, recompute the leaves against the artifacts, rebuild the root, and verify the signature against our published public key without ever talking to Felarity. The verifier endpoint at /api/verify performs the same check in-browser for convenience, but the math does not require it. If a single transcript line, audio segment, or council finding has been altered after sealing, verification fails.
30 industry councils
Every workspace selects an industry on creation. The council for that industry is a named panel of analyst personas — each one a specialist role that critiques the same content from a different angle. The same paragraph of meeting transcript that looks unremarkable to a generalist looks load-bearing to a regulatory affairs lead, and dangerous to an ethics officer. The council surfaces all of those readings at once and the synthesis layer reconciles them with attribution.
Exports and shareable reports
Reports export to PDF with the full attestation appendix bound in, to Markdown for ingestion into matter-management or knowledge-base systems, and to JSON for programmatic downstream use. Each export carries the same signature as the source record — exporting does not break verifiability. For sharing outside the workspace, generate a password-protected read-only link with a hard expiry; revoke at any time from the workspace admin panel. Share-link access is logged per view with IP and user-agent in the workspace audit trail.
API and integrations
The REST API uses key-scoped Bearer tokens prefixed flrt_, each one issued against a specific workspace with its own rate limit and its own audit trail. Every authenticated call writes a row to the per-key event log, which is queryable from the admin surface and exportable to your SIEM. The full OpenAPI specification is at /api/openapi.json, and webhooks fire on report.ready, contradiction.flagged, and verification.failed so your downstream systems can pull the attested report as soon as the pipeline completes.
SAML SSO covers Okta, Azure AD, and Google Workspace on Professional and above. SCIM 2.0 user provisioning is available on the Enterprise tier. Slack notifications post contradiction alerts and report-ready events into the channel of your choice, with the share link and the attestation fingerprint inline.
Self-hosted and dedicated tenancy
Enterprise customers can run Felarity on dedicated infrastructure — a single-tenant VPC managed by our team, or fully on-premise inside your own environment. Both options support bring-your-own-key LLM, so the council, the synthesis layer, and the NLI scoring can route through the model your security team has already approved rather than the one we ship by default. No meeting content transits a shared model. Dedicated and on-prem deployments are bookable through sales.
Walk through the product.
Bring a sample meeting or use ours. We will run the live phase and the full forensic pipeline against it on a call, and you keep the attested report.