Stable source identity and canonical URL.
Products
OBJECTProduct
PUBLIC STATUSACTIVE BUILD · WORKING MVP CORE
FAMILYKNOWLEDGE EXPLORATION SOFTWARE
Projectr
Turn long-form sources into knowledge you can navigate and keep.
Projectr is a durable knowledge-exploration product. YouTube Knowledge Explorer is its current working implementation: preserve the source, keep timestamps, search evidence, follow concepts, and ask questions that point back to the original material.
CURRENT PRODUCT BOUNDARYThe current YouTube implementation has a working core for source parsing, transcripts, outlines, search, evidence-linked answers, persistence, and interchange. Repeated use, retention, pricing, and product-market fit remain unestablished.
ILLUSTRATIVE SOURCE · 1:07:14One long video. Multiple lawful projections.
SYNTHETIC PRODUCT DEMOMaking a long source navigable without losing the source34:12 · Search should return evidence
TIMESTAMPED EVIDENCE5 demo segments
WHAT PROJECTR DOES
Make a long source easier to search, revisit, and understand without hiding the original.
Long-form video carries explanation, argument, demonstration, and context, but precise retrieval is expensive. The Explorer turns one source into a navigable knowledge object without losing the path back to the original moment.The useful unit is not just a feed item. It is a source you can search, navigate, save, revisit, and trace back to the original moment.
Authorized captions or an imported VTT/SRT source.
Normalized evidence units with start and end times.
A deterministic topic / outline projection over the source.
Portable enrichment and evidence-linked retrieval.
Answers cite supporting source segments or say that the source does not contain enough evidence.
A portable package containing the source, transcript, outline, concepts, and exploration state.
THE CURRENT YOUTUBE IMPLEMENTATION
A 90-minute video should be easier to work with than a scrub bar.
Outline, transcript, concepts, search, and source navigation are different views over the same source. None of them replaces the original video.Jump between sections without flattening a 90-minute source into one summary.
- 06:18Start with the source
- 18:42Structure without replacement
- 34:12Search should return evidence
Find words, phrases, and concepts, then open the exact timestamp that contains the evidence.
searchrepresentation34:12 · 41:08 · 58:27
Concepts link back to the supporting transcript and topics rather than becoming detached tags.
sourceevidencetimestampoutlineadmissibility
Hosted synthesis is optional; the product can fall back to deterministic extractive answers.
Claims remain connected to the source segments that support them.
34:1258:27
ASK THE SOURCE
Ask a question and keep the answer tied to the moments that support it.
A model may help write the answer, but the supporting moments must come from the loaded source. If the source does not support the answer, the product should say so.ANSWER FROM SOURCE EVIDENCE2 supported claims
- Timestamped segments let navigation and search point back to the original moment rather than detaching an answer from its source.Evidence · 34:12
- Source identity, transcript structure, and evidence references can survive a change in the model or interface used to explore them.Evidence · 58:27
Every answered claim names one or more transcript segment IDs.
Displayed source excerpts are reconstructed from the loaded transcript, not authored by a model.
Evidence remains jumpable back to the source moment.
Topic and concept references must resolve inside the current exploration.
When the source does not support an answer, the valid result is to say so.
HOW IT IS BUILT
Your saved knowledge should not be trapped in one interface or AI provider.
The useful knowledge object should survive changes in interface, storage, hosting, or AI provider.
Source, transcript, map, concepts, answers, saved exploration.
Parsing, normalization, outline derivation, admissibility, package validation.
Explore, search, ask, save, load, import, export.
Metadata, transcript, enrichment, search, answering, persistence.
YouTube APIs, browser storage, local files, hosted or deterministic answering.
YOUTUBE / CAPTIONS→PORTABLE PROJECTR CORE→BROWSER / FILE / DATABASE→OPTIONAL MODEL ADAPTER→PORTABLE EXPLORATION
CURRENT PRODUCT STATE
The core works. The next question is whether people choose to keep using it.
Build something useful first. Treat repeated voluntary use as the next evidence.
- YouTube URL parsing and portable source metadata
- Authorized caption acquisition plus VTT/SRT import
- Transcript normalization and exact / concept-linked search
- Deterministic outline and concept enrichment
- Evidence-bound deterministic and hosted answering
- Browser-local persistence and versioned import / export
- Polish the current YouTube product experience
- Run real source and user sessions
- Measure repeated voluntary use and failure modes
- Then test willingness to pay and packaging
Not every YouTube video exposes captions that can be lawfully acquired.
A source-linked answer is not automatically a true statement about the world beyond that source.
The current concept layer is useful enrichment, not a complete ontology of the video.
The product does not claim to replace watching, reading, or interpreting the original source.
Repeated use, retention, willingness to pay, market demand, and product-market fit remain unestablished.
PRODUCT DIRECTION