VizhiCast: AI that turns event footage into a personal highlight reel for every guest

The business problem

At a wedding, concert or college fest, cameras roll for hours, but every guest really wants just one thing: the moments they're in.

Today that means someone scrubbing through hours of footage from several cameras, finding each person, cutting clips and stitching them together. Doing it for one guest takes hours, and for hundreds of guests it never happens. Photographers and studios end up delivering one long event film that most guests never watch, and they miss out on a product people would happily share and pay for: their own personal reel.

I wanted to make that automatic: upload the footage once, and every guest gets a highlight reel of themselves, cut to music, in minutes.

Why existing solutions failed

• Manual editing doesn't scale. A good editor can make a handful of personal reels a day, not hundreds per event.
• Photo face-search tools (common at weddings and marathons) only find guests in still photos, not in video, and they don't produce anything shareable.
• Generic auto-edit apps make one highlight video of the whole event. They don't know who anyone is, so they can't make it about a specific guest.
• Cloud AI video services are expensive per minute of footage, send guests' faces to third parties, and give organisers little control over accuracy.

The solution

VizhiCast is my own product: an AI pipeline plus an organiser dashboard that turns raw event footage into personal, beat-synced highlight reels.

1. The organiser creates an event and uploads footage from one or more cameras.
2. Each guest provides one reference photo, with consent recorded first.
3. The AI finds every moment that guest appears, cuts those moments out and merges them into a reel.
4. The reel is synced to music, with optional beat-timed effects, ready to share.

Face matching you can trust

Lookalikes are the main way automatic reels go wrong, so matching is deliberately strict:

• Footage is downscaled and sampled for speed, then every detected face is compared with the guest's reference encoding using dlib.
• A match must fall within a tunable tolerance (stricter than the library default), clearly beat the second-best match, and repeat across several frames. Single-frame flukes are ignored.
• Matches are grouped into presence windows with natural lead-in and lead-out, so reels contain full moments rather than choppy fragments.
• Organisers watch live 0–100% progress with a count of matches found so far, and every run writes its own log for troubleshooting.

Privacy built in

Faces are sensitive data, so the flow is consent-first:

• Guests tick a consent box before their photo is processed, and the consent is stored with the event.
• The photo is converted to a face encoding on the server and the original is deleted immediately.
• Encodings can be removed at any time from Manage Faces.
• Everything runs on open-source models on my own server. No faces are sent to third-party AI APIs.

Beat-synced reels

Upload any music track and librosa detects its beats. FFmpeg then replaces the reel's audio and can add beat-timed flash, pulse, zoom or beat-cut effects. Organisers preview the result before applying it.

Follow Me: multi-camera reels

With two or more cameras, Follow Me scans every angle, scores how directly the guest is facing each camera, and builds one chronological reel that always cuts to the camera they're looking at, like a multi-camera edit done for each guest.

AI Studio: describe the reel you want

Guests or organisers can type a prompt like "exciting dance moments on the stage", pick a style (Cinematic, Energetic, Emotional, Documentary) and choose a length from 15 to 60 seconds:

• A vision-language model (Moondream) captions a frame from each candidate clip.
• A sentence-embedding model (Sentence-BERT) ranks the clips by how well they match the prompt.
• The best clips fill the chosen length and are merged into the final reel.

The same flow powers a guest portal: guests enter their event, add a selfie and an optional prompt, and get their own reel without the organiser doing anything.

Live capture from real cameras

A small camera agent script records from DSLR/mirrorless cameras (via HDMI capture), GoPros and IP cameras (RTSP), or OBS. It uploads footage in 30-second chunks as the event happens, retrying failed uploads and keeping any chunk that never makes it, so processing can start before the event ends.

Key features

• Personal highlight reels from raw event footage, one per guest
• One reference photo per guest, with no cap on footage length
• Strict face matching: tolerance, best-versus-runner-up margin and repeat-frame checks
• Full moments, not fragments, through presence windows with lead-in and lead-out
• Live progress and per-run logs
• Consent-first privacy: photos become encodings and are deleted immediately
• Beat-synced music with flash, pulse, zoom and beat-cut effects
• Follow Me multi-camera cutting
• AI Studio: prompt-to-reel with Moondream and Sentence-BERT
• Guest portal for self-service reels
• Camera agent for DSLR, GoPro, RTSP and OBS chunked uploads

Technical architecture

• Admin panel: Next.js 15, React 19 and Tailwind CSS 4, with Prisma for events, face profiles and consent records. It handles uploads, video streaming, FFmpeg thumbnails, reel management and a protected admin login.
• AI service: Python/Flask, packaged with Docker (FFmpeg, dlib, OpenCV and audio libraries included). Endpoints cover reel generation per event, live progress, face encoding, audio beat-sync and AI Studio jobs. Heavy models load lazily, so startup stays fast.
• Shared workspace: the admin panel and the AI service read and write the same input/ and output/ folders, so no media is copied between services.
• Hosting plan: AWS EC2 for both services, S3 + CloudFront for reels, and a GPU instance for AI Studio when demand grows. I documented a cost breakdown by tier.

Challenges

• The wrong person in a reel. Early runs occasionally mixed up lookalikes. Adding a best-versus-runner-up margin, a minimum match count and a minimum gap between matches fixed this, and per-run logs made every decision traceable.
• Choppy reels. Cutting a fixed 3-second clip around each match produced jittery results. Switching to presence windows, continuous spans where the guest is on screen, made reels feel edited by a person.
• Long footage. Hour-long event videos are slow to scan. Downscaling, frame striding and removing the frame cap keep full-length footage practical, and live progress keeps organisers informed.
• Multi-camera timing. Follow Me needs to know which camera to show at each moment. Scoring how directly the guest faces each camera, bucket by bucket in time, gives natural camera switches.
• dlib on Windows and Docker. Installing face recognition usually means compiling dlib. Using prebuilt wheels locally and a Docker image for servers made setup repeatable.

The outcome

• A working end-to-end product: upload footage, add a guest photo, and get a personal, beat-synced reel, with live progress, consent handling and multi-camera support.
• Hours of editing become minutes of processing per guest, with no editor in the loop.
• A new product for event businesses: a personal reel for every guest is something studios can sell and guests want to share.
• Next: a guest mobile app (in progress, built with Expo) for joining events, scanning a selfie and receiving reels on the phone.

What this means for other businesses

Any business that records lots of video of lots of people, such as wedding studios, event companies, schools, sports clubs, marathons and conferences, is sitting on content its audience would love, if only it could be found and cut per person.

• Face recognition and smart editing can turn one long recording into hundreds of personal, shareable videos.
• Open-source AI on your own servers keeps costs predictable and sensitive data private.
• Personal content sells: it's an upsell for studios and a reason for guests to share your brand.

If you have hours of footage and an audience that wants their moments, I can help you build the pipeline that delivers them.