Video, audio and captions turned into searchable, attributed knowledge
From a content pipeline shipping to ten platforms behind a human gate, to a table-tennis coaching index and a film-dialogue search engine — one underlying capability, five projects, honestly labelled.
The problem
Recorded material is where the answers already are, and the hardest place to look: hours of video, audio or captions hold the one moment that settles a question, and nothing indexes it. The same gap runs the other way — turning a single idea into the many shapes each platform wants is slow, repetitive work.
What the portfolio proves
Transcript and caption pipelines, taxonomy and lexicon design, pose estimation on video, always-attributed retrieval, and a human-gated content production pipeline that turns one idea into platform-specific packages.
How the projects fit together
Content Creator (In use) is the pipeline actually in regular use. It carries an idea through qualification, scripting and drafting, then packages it for ten platforms (about 194 passing tests across 10 adapters), gated by human review before anything ships. FilmyGyan (Prototype) turns Indian-cinema dialogue into a tagged, always-attributed knowledge base and persona chat (about 163 passing tests). KinoAI (Prototype) tags coaching-video transcripts against a stroke/skill taxonomy (239 nodes, 1,797 lexicon terms, 39,779 tagged cues) — the knowledge layer for a planned camera-only coaching product, which remains a plan, not a build. From TT Global Studio (Parked) is a dormant worktree of KinoAI that extracts per-stroke body-position templates from the same coaching video; mention it beside KinoAI, not as a separate product. DashCam (In use) sits here as secondary evidence of the same footage-understanding capability, described in text only per its publishing block.
What I can do here
Media, marketing and content teams evaluating this capability, and coaches or academies wanting transcript- or footage-search, are pointed at the same Services offer that scopes a pilot around one platform or one video corpus first.
How mature this is
Only the content pipeline is in regular use; the sports-coaching work is a knowledge layer plus a parked pose-analysis prototype, not a finished coaching product. Source video and dialogue are third-party material, so no frame, clip or quoted dialogue is published anywhere on the site.
Mapped projects
Content Creator
In use
Running today and in regular use; it is not an open public service you can sign up for.
Turns one idea into publish-ready content packages for ten platforms, with a human approval gate.
Producing consistent, on-brand content across many social platforms from a single idea is slow and error-prone.
A git-versioned pipeline that carries an idea through qualification, scripting and drafting, then packages it per platform (YouTube, Shorts, Instagram, TikTok, X, LinkedIn and more), gated by human review before anything ships.
Ten platform adapters, each with its own spec and prompt.
The human-approval gate sits before anything ships, not after, so nothing is published automatically.
Python
Claude Code skills and a platform-writer subagent
optional local media tooling
DashCam
In use
Running today and in regular use; it is not an open public service you can sign up for.
A pipeline that ranks, transcribes, speaker-labels and mood-tags dashcam footage, so a moment can be found by the words spoken. A companion Android app does the same recall entirely on the phone. No clip, frame or still is shown here: the project's own policy blocks publishing footage until face and plate blurring and speaker consent are in place.
FilmyGyan
Prototype
The core works end to end and can be demonstrated; it is not production-hardened or generally available.
Turns Indian-cinema dialogue into a searchable, always-attributed knowledge base and persona chat.
Finding "the scene where a character says the thing that fits this situation" is otherwise a manual, memory-dependent search through films.
Ingests film captions into a tagged dialogue knowledge base, powers situation-based search, and runs a persona chat that always attributes its quotes back to the source film.
The always-attribute-the-source behaviour is the point, given the dialogue is third-party copyrighted material.
The production tier caps quote length rather than storing full text, so the site describes the search and attribution capability and reproduces no quoted dialogue.
Python
SQLite
TF-IDF retrieval
optional AI-assisted tagging
From TT Global Studio
Parked
Built far enough to prove the idea, then paused; it is not in active development.
A dormant worktree of KinoAI. It finds the moments in coaching video where a coach demonstrates a stroke, then aggregates a per-stroke, joint-angle body-position template from them. It is mentioned beside KinoAI because it shares that knowledge base and taxonomy, not as a separate product.
KinoAI
Prototype
The core works end to end and can be demonstrated; it is not production-hardened or generally available.
An index of exactly which coaching video teaches a given table-tennis skill, built toward a camera-only AI coach.
There is a huge amount of table-tennis coaching video online, but no way to find the exact clip that teaches a specific skill or fault fix.
Ingests public coaching-video transcripts, tags them against a stroke/skill taxonomy, and builds a searchable index of teaching moments — the knowledge layer for a planned camera-only AI coaching product.
A 239-node taxonomy with 1,797 lexicon terms tags 39,779 cues, covering 366 of 459 known videos.
The camera-only coaching product itself remains a plan, not a build, and the copy says so.
