How I AI: Anish Acharya's 3 Creative AI Workflows for Music Videos, Book Cataloging, and Personal Finance

AI-generated grunge concert footage, a bookshelf catalog built from a quick phone video, and a browser agent inspecting a live investment account all happen in the same conversation. In this episode of How I AI, Anish Acharya, a general partner at Andreessen Horowitz focused on consumer investing and AI-native products, demonstrates three personal workflows: assembling a synthetic music-video experiment, cataloging books from a phone video, and using a browser agent to inspect portfolio information.

The interesting part is not just the tools. It is how Anish approaches them. Each workflow starts with a personal curiosity, then turns into a fast prototype using multimodal models, lightweight editing tools, and a willingness to iterate until the aesthetic or output feels right.

Building a synthetic Tiny Desk-style music video

Anish started with a synthetic image of Kurt Cobain performing in a Tiny Desk-inspired setting, then kept refining the framing and mood until it felt closer to a grainy 1990s live performance than a polished AI render.

Generating the concert still in GPT-4o

Anish used GPT-4o to generate an image of Kurt Cobain performing a Tiny Desk-style concert. He likes GPT-4o for image work because it follows aesthetic instructions closely while still leaving room for unexpected results. During the demo, he pointed out details like the correctly rendered NPR logo and the cardigan styling that made the image feel grounded in a specific era rather than generic grunge nostalgia.

He removed the guitar in a later version to create a tighter vocal-performance shot. That image became the anchor frame for the rest of the workflow.

Animating the still image with Hedra

He brought the still image into Hedra, which generates motion from a static frame while synchronizing facial movement to uploaded audio. Anish described it as one of the fastest ways to combine frame-to-video generation and lip sync in a single workflow. He also mentioned Sync Labs as another tool he likes for similar work.

Preparing and syncing audio in Adobe Audition

Anish downloaded a live Nirvana performance clip, opened it in Adobe Audition, trimmed silence from the beginning, and isolated a short excerpt to sync against the generated performance. Adobe Audition let him scrub through the waveform while previewing the video reference, making it easier to choose a segment that matched the pacing and emotional tone of the synthetic performance. Anyone repeating the workflow should use audio they own, license, or are otherwise permitted to download and transform.

Separating vocals with Demucs

He used Demucs to split vocals from instrumentation before assembling the final sequence. Anish described this capability as one of the breakthroughs that originally pulled him into AI music workflows because older production pipelines treated fully mixed tracks as effectively locked. Source separation does not create rights to reuse the recording.

Creating the final music video aesthetic

Anish assembled the final edit in Kapwing and mixed in short atmospheric clips generated with Veo 3. The first generations missed the tone completely, looking more like polished synth-pop videos than grimy Pacific Northwest footage. He used GPT-4o to refine the prompts toward a camcorder-heavy, dystopian 1990s Seattle aesthetic, then kept iterating until the clips developed the washed-out textures and awkward crowd energy he wanted. The AI music-video workflow lays out the same path from a generated anchor image through animation, audio preparation, lip sync, and final assembly.

The finished piece looked intentionally synthetic rather than disguised as archival footage. You could still spot artifacts in crowd motion and object handling, but the emotional tone, wardrobe, and camera movement were convincing enough that I immediately started brainstorming her own fake trailers and music videos. Work using a real person's likeness should be labeled and reviewed for applicable publicity, platform, and deception rules before distribution.

Turning a bookshelf video into a searchable catalog

Anish's second workflow uses Gemini Flash to extract book titles and authors from a short phone video instead of requiring manual inventory entry.

Generating the app in Google AI Studio

Inside Google AI Studio, Anish prompted the system to create an app that accepts a video of a person flipping through a book collection and extracts every visible title and author. He explicitly suggested a workflow where the app first samples frames from the video, then sends those frames through a vision model for analysis.

Extracting titles from sampled frames

The app processes the uploaded bookshelf video frame by frame, identifies distinct books, and asks Gemini Flash to recognize titles and authors. During the demo, the system successfully captured a mixed stack of books after only a few seconds of recording. It also surfaced some duplicates and imperfect detections, which reinforced the practical limitation here: OCR and visual identification are useful accelerators, not trustworthy inventory systems without review.

Deploying the prototype with Cloud Run

Anish deployed the prototype through Cloud Run so it could be shared with other people instead of remaining a local experiment. He noted that getting a personal prototype working took minutes, while making it usable for others took much longer because deployment introduces authentication, storage, cost, and moderation concerns. A shareable deployment also needs upload limits, retention rules, cost controls, and a clear policy for videos that may reveal a person's home or possessions. The Gemini Flash book-catalog workflow shows the prompt, frame-sampling approach, correction pass, and deployment checks together.

Using a browser agent to inspect portfolio data

Anish also demonstrated Comet, Perplexity's browser agent at the time of recording, using his own Robinhood account as the test environment.

Researching investments with Comet

A browser agent can navigate websites, inspect information, and answer questions using the same permissions as the signed-in user. In practice, Anish used Comet to summarize portfolio performance, compare holdings, and surface additional investment ideas without manually downloading spreadsheets or clicking through dashboards.

The convenience is obvious, but this workflow also has the sharpest operational boundary of the episode because the agent inherits access to whatever the browser session can reach. Anish limited the workflow to research and exploration rather than account actions. Anyone adapting it should begin with read-only access, exclude trading and money movement, protect credentials and account data, verify every figure against the brokerage, and treat generated analysis as research rather than financial advice. The Comet personal-finance workflow makes those permissions and verification steps explicit.

Three workflows, three very different trust boundaries

What makes these workflows useful is not that the AI fully replaces human judgment. The music-video workflow still requires taste, editing choices, and careful handling of likeness rights. The book catalog still needs correction and privacy review. The browser-agent workflow only works safely if account permissions stay tightly constrained and every important number gets verified independently. In every case, the AI output is a draft artifact that speeds up exploration rather than a final authority.

The strongest pattern worth copying from this episode is the combination of multimodal input with lightweight iteration. Anish consistently starts from messy real-world material like videos, songs, screenshots, or browser sessions instead of perfectly structured data. That approach works especially well for creative experiments and personal utilities where speed matters more than precision. The part that still requires human judgment is deciding what should remain synthetic, what needs verification, and where convenience crosses into risk.

Anish's prompting advice follows the same experimental logic. For creative work, he starts with short prompts that leave the model room to surprise him, then narrows the direction after he has something concrete to react to. When a coding branch or generation keeps collapsing, he abandons it and starts again instead of treating the model's effort as his own sunk cost. The practical lesson is to preserve exploration early, but become ruthless about restarting once the output stops teaching you anything.