How I AI: Tim McAleer's AI Workflows for Documentary Filmmaking at Florentine Films

In this episode of How I AI, Florentine Films producer Tim McAleer shows how he uses AI to organize documentary archives, pair field research with its source notes, and extract text from difficult historical documents.

A Ken Burns documentary can draw on an enormous archive before a single frame reaches the screen. For the Muhammad Ali series, the team gathered more than 20,000 still images and hundreds of hours of footage. Every asset needed descriptions, tags, and sourcing information. Tim knew the cost of that work firsthand: "it used to be my job, so I can tell you firsthand, not my favorite part."

Tim builds narrow tools for the operational work around filmmaking rather than generating the films themselves. His systems help researchers catalog, find, and evaluate source material while the editorial decisions remain with the team.

The three workflows cover a media-analysis service, an iOS capture app called Flip Flop, and a Mac utility called OCR Party. Together they show how a team can add structure at capture time, carry source context into analysis, and keep AI output reviewable.

Build an AI-powered database for archival assets

How to Create a Custom Mac App for Precision AI-Powered OCR: Build a focused Mac utility for selecting one article, signature, or passage inside a difficult historical scan. OCR Party crops the exact region, offers local or AI OCR, returns editable text, and records where the crop came from.

How to Streamline Archival Research with a Custom AI-Powered iOS App: Build a field research app that captures the front and back of an archival photo as a linked pair, generates a description and transcription immediately, embeds both into image metadata, and gives every file a predictable name.

How to Build an AI-Powered Database for Archival Media Assets: Turn archival images, video, and audio into searchable database records. Tim combines embedded metadata, web research, visual captions, Whisper transcripts, and fused embeddings so researchers can find assets by meaning, not only exact keywords.

Documentary post-production involves more than storing files. Each image, video clip, and audio recording needs technical metadata, a useful description, and a traceable source. Tim built a pipeline that turns incoming media into structured records without requiring a researcher to type every field.

move from manual prompts to an automated script

Tim began by testing image uploads in ChatGPT with archival photos. The visual descriptions were promising, but a chat window could not handle production volume. He used Cursor to turn the experiment into a Python script that called the OpenAI API.

He dictated the first request into Cursor with Superwhisper:

"Write me a script that submits the jpeg at the root of this workspace to open ai for description. I want just a general visual description of what we can see in the image. Uh, any API credentials you need are in a text file at the root of the folder."

The first result was descriptive but uncertain: "This image depicts a small rural main street from what appears to be the mid 20th century." That language is not sufficient for a documentary record that will be fact-checked.

ground descriptions in embedded metadata

Archival files, including material from the Library of Congress, can contain embedded photographer, date, location, and accession information. Tim changed the script so the model received that metadata alongside the pixels instead of inferring the record from the image alone.

"I want you to add a step to this script. I wanna scrape any available metadata from the file first and append that to the prompt."

The script extracted the file metadata and passed it into the prompt with the image. The resulting description became more specific:

"The image shows a street scene on the main street of Cascade, Idaho. Captured in 1941 by photographer Russell Lee."

Metadata provides evidence, not automatic truth. Tim's pipeline can also retrieve context from source pages, and the production team still verifies information used in a film.

process video and audio together

Tim expanded the pipeline so a video could be represented by timed visual and audio events:

1. Frame sampling: The service captures a still every five seconds rather than sending every frame to a model.
2. Frame captioning: Each sampled frame goes to a lower-cost model such as GPT-5 nano for a short visual caption.
3. Audio transcription: Whisper transcribes the audio in timed chunks that correspond to the sampled frames.
4. Reasoning and summarization: A reasoning model receives the frame captions, timed transcript, and available metadata, then produces a description of the clip.

I noted that Gemini's video models can identify useful still frames automatically. Tim treated that as a possible future addition rather than part of the demonstrated pipeline.

add semantic search with vector embeddings

Descriptions made assets readable, but exact keyword search still limited discovery. A researcher searching for "puppy," for example, might miss an image described as "dog."

Tim added a multimodal semantic-search layer for each asset:

• The image thumbnail becomes a vector embedding through the open-source CLIP model.
• The text description becomes a second embedding through an OpenAI text model.
• The service combines the image and text representations for retrieval.

Editors can select an image and retrieve visually or semantically related material from the collection. The result is another way into a large archive, not a substitute for reviewing the underlying records.

Build Flip Flop for field research

Researchers often photograph folders of documents in physical archives. A typical camera roll separates the front of an item from handwritten notes on its reverse and leaves thousands of files to reconcile later.

Tim described the screens and capture flow he wanted, then used ChatGPT to draft a product requirements document. Claude Code did not build the entire app in one pass, but it did produce the SwiftUI interface in one pass. Tim continued working on the app from there.

Flip Flop gives the capture process a consistent structure:

1. Create a collection: Start a collection for the folder or box being reviewed.
2. Capture front and back: Photograph the front, then use the "flop" action for the reverse side.
3. Run AI processing: Send the paired images to the OpenAI API after capture.

• Describe the front image.
• Transcribe notes or other writing on the back image.

1. Embed the context: Write the description, transcription, and collection details into the image file's EXIF metadata.

The naming scheme keeps fronts and backs together, while the embedded metadata travels with each file. Two colleagues returned from one archive trip with 1,400 images. Tim did not attribute that volume solely to Flip Flop, but said the app made processing the material easier after they returned.

Use OCR Party for selected document regions

Full-page OCR can struggle with old newspapers, faded scans, multiple columns, or handwritten letters. Tim built a Mac menu bar app for cases where a researcher needs text from one particular region.

OCR Party takes its name from the action it supports: running OCR on part of an image.

The workflow keeps the source region attached to the result:

1. Open the source: Load the document image in OCR Party.
2. Select a region: Draw a box around the article, paragraph, or handwritten name that matters.

1. Choose an OCR method: Submit the crop either to macOS Vision or to an AI API. AI can reconstruct plausible text across creases, faded areas, and ink marks, but those reconstructions may be wrong.
2. Review the output: Copy the extracted text and retain the crop coordinates so an editor can return to the original. Tim explicitly noted that any text used in a film would be fact-checked later.

I connected the tool to the difficult 18th-century cursive her mother encounters in genealogical research. Selecting one region at a time can make a dense source easier to investigate without pretending the transcription is authoritative.

Where Tim's approach works

Each tool addresses a specific production bottleneck: cataloging media, capturing paired archive images, or transcribing a selected document region. The narrow scope makes the output easier to inspect and fit into an existing process.

The larger payoff comes from consistent structure. Researchers spend less time moving text between systems, editors gain new ways to search the archive, and every result can still point back to its source.

A useful starting point is one repetitive task with a clear input, output, and review step. Tim's work shows that a small internal service or single-purpose utility can be valuable without trying to automate the creative judgment at the center of documentary filmmaking.