Summary
Three to six paragraphs. The shape of the conversation, what the student was working through, where it landed. Pinned to the session row.
Bring a session in — a Zoom recording, a transcript upload, or both. Observatory runs a structured synthesis pass and returns the room you sit in to review it: a player where every highlight, chapter break, flag, and suggested tag is timecode-keyed back to the moment that produced it. Searchable. Annotated. Signed-off. Click a highlight, the player seeks to it. Mark a clip, render a reel, hand the whole thing to a parent or a reviewer. Built for the program we run; productized inside an Avira install.
Most teams have meetings, recordings, transcripts piling up unprocessed. Observatory takes one — a Zoom recording it pulled itself, or a video and transcript you uploaded — and runs a synthesis pass that produces five artefacts pinned to the session: a summary, a set of highlights with timestamps, chapter breaks, severity-rated flags against categories you defined, and suggested tags. Every artefact is keyed to a moment in the recording. The player jumps to it.
Around that core, the room: timestamped notes, a clip composer that renders to MP4 in the background, a reel builder that stitches clips into a single sequence, a tag editor with operator-curated suggestions, role assignment for the people in the session (student, parent, facilitator, observer), and a ZIP export with the metadata, transcript, summary, and source artefacts when you need to hand the whole thing to someone outside the system.
The synthesis itself runs the cache discipline from SoQ LLM. The transcript is the long, stable prefix; each pass is a small varied suffix. Five passes against the same transcript pay for the prefix once. Every run is ledgered with model, prompt tokens, cached tokens, and outcome.
Sample: a NorthStar dialogue, 47 minutes, four people in the room. The synthesis pass has run; the operator is reviewing.
The session opens with the student framing the difficulty as “I don’t know what I want to make,”a clean uncertainty rather than a defeated one. The dialogue reorients away from career questions and toward what has been alive recently — a recurring conversation about local water, a pull toward textile work, an unfinished story.
The pivot lands at 14:22 when the facilitator names the pattern aloud and the student stops mid-sentence. From there: three concrete leads emerge by the close. The next session has been scheduled for two weeks out, returning to the textile thread first.
Three to six paragraphs. The shape of the conversation, what the student was working through, where it landed. Pinned to the session row.
Timecode-keyed moments worth returning to. Each is a click-to-seek anchor on the player. Out-of-range timestamps are clamped, never silently dropped.
The session split into named segments. Renders as a strip beneath the scrub bar; the operator can rename or merge before sign-off.
Severity-rated against categories your operator defined — “parent question to revisit,” “safeguarding signal,” whatever your room actually needs to surface. Categories are configurable from settings.
Two to five tag candidates the operator can accept, edit, or ignore. Existing tag values bias the suggestions; nothing is auto-applied.
Every pass is ledgered: which model ran, prompt tokens, cached tokens, completion tokens, outcome. You can answer the question “which model produced this insight, on what date” without guessing.
Clips render in the background while you keep working; reels stitch the rendered clips into a single sequence. The operator never leaves the session.
A real synthesis run is auditable end to end. The cache discipline shows up in the third column: pass one seeds, passes two through five hit. Sample run shape; your numbers will look like this.
| Pass | Model | Prompt | Cached | Completion | Status |
|---|---|---|---|---|---|
| Summary | claude-sonnet-4-6 | 16,210 | — seeds | 1,142 | ok |
| Highlights | claude-sonnet-4-6 | 16,225 | 16,180 | 1,860 | ok |
| Chapters | claude-sonnet-4-6 | 16,225 | 16,180 | 720 | ok |
| Flags | claude-sonnet-4-6 | 16,250 | 16,180 | 480 | ok |
| Suggested tags | claude-sonnet-4-6 | 16,225 | 16,180 | 220 | ok |
It is not a generic LLM tracer. If you want OpenTelemetry over every model call your stack makes, that is a different category and we are not it. Observatory is a session room — a recording, a transcript, the structured insight pass we built around them.
It is not session intelligence for sales calls. Built for coaching, teaching, and review sessions where one human is responsible for what the room contains afterwards. Outbound-call analytics are a different shape; we will tell you on the first call if your case is the wrong fit.
It is not a standalone product you self-serve. Observatory ships inside an Avira install, configured to your substrate — your students, your sessions, your tag vocabulary, your flag categories. The shape of the engagement, and what we will not do, lives on that page.
Observatory ships inside an Avira install. Write to us with a sample session you would want this run on, and we will reply within three working days, from a person.