# Agentic AI Tech Solutions: from portfolio film to focused hero cut

Production record · 22 September 2026 · AIC Mastermind FilmCraft Studio Community Edition

A company film became a useful test of the workflow precisely because its first revisions were not good enough. The long version showed range but asked viewers to remember seven stories. The first hero edit shortened it, then a readability repair removed too much demonstration. The third edit restores visible state changes while retaining speech-based captions.

## Watch and compare

- Original portfolio: **176.96 seconds**, six products plus the internal workspace. Preserved unchanged.
- Speech-aligned hero: **83.99 seconds**, four products across three connected stories. Preserved as a revision example, not the selected current cut.
- Proof-led hero, v3: **84.27 seconds**, the same four products with illustrative state changes and a restored conversational synthesis.

The files in this package are the original exports. The 848×480 figures in the external review describe the copies that reviewer received; those copies were not available for matching. Do not attribute those measurements to the original masters.

## Brief and audience

Vikram Raviprolu commissioned a marketing film positioning Agentic as a product company. The initial request allowed about 120 seconds; later instructions explicitly removed the time limit and prioritized natural conversation. The working audience was business owners, operators, content teams and prospective product clients. The desired action was to explore Agentic's products or describe a business process that needs improvement.

The common promise was “Start with real work.” The film was not permitted to invent measured productivity gains, imply all products share a live integration, expose customer records, or present illustrative interfaces as authenticated product capture. Website/presentation use was the working destination; no social upload or platform transcode is claimed by this record. Vikram is the release decision owner. A fixed total production spend ceiling and a delivery deadline were not recorded in the recovered production documents.

## Why these products

| Story | Product role | What v3 shows | Evidence boundary |
|---|---|---|---|
| Research to editorial control | GapWright + Content OS | Example evidence becomes a chosen content opportunity; a brief becomes two formats and an approved exact revision | Fictional editorial demonstration; not a live integration |
| Check the answer | VortexHR | A handbook question leads to a sample answer and highlighted source clause; missing evidence stays unresolved | Fictional policy; not a live retrieval or legal guidance |
| Know the next action | Rank-in-my-Pocket | Fictional quote moves to illustrative order and readiness action | Reconstructed demo, not a transaction in the production app; support labelled coming soon |

Reelcraft, VyapaarOS and the internal workspace remain in the long portfolio film. They were removed from this hero for focus, not withdrawn from the portfolio. Separate VyapaarOS TEST screenshots were collected later; they are not used here and do not prove an order conversion. The Rank production workspace was inspected read-only. Its customer data was not copied into this film.

## Version history and repairs

| Version | What worked | What failed | Repair |
|---|---|---|---|
| Full portfolio | Coherent identity and product stories; conversational voice | Long holds, dense details, too many stories for a hero | Select four products and shorten the route |
| Initial hero | Shorter, clearer structure | Fixed-time cards competed with speech | Recover actual spoken-word timings |
| Speech-aligned hero | Large, synchronized phrases | Became a transcript; product demonstration disappeared | Separate proof area from subtitle band |
| Proof-led v3 | Visible illustrative state changes, retained voice timing, clear provenance labels | Final audience response remains to be collected | Keep original films and source records so future edits remain reversible |

The v3 repair is not evidence that each application performed these actions during capture. It is a clearer, explicitly illustrative explanation of the workflow. There is no customer testimonial or measured business outcome in the film.

## Canon and layout

The official source logo images remain unmodified. The composition uses Arial, a dark teal field (#091b22), off-white text (#f5f5e9), lime (#d9ff82), and dark panels (#15353d). The full-width version uses 80px side margins; its main proof text is 68px and captions 59px. Required evidence labels are 52px. The 9:16 version is recomposed at 1080×1920 with 55px side margins, 62px main proof text, 58px captions and protected top/bottom bands. It is not a centre crop.

The headline identifies the product; the central area shows one state or decision; the lower band carries the spoken phrase. Colour changes identify a selected opportunity, approved revision, source clause or order state. No word-by-word bouncing captions or rapid decorative transitions are used. Phone-size contact sheets form part of the QC package; accessibility is not inferred from font size alone.

## Script and voice

The long version contains 22 retained dialogue takes, voiced by Faye (business owner), Ainsley (Agentic guide) and Isla (creator), using Higgsfield's seed_audio presets. These are synthetic voices, not Maple or recordings of named employees. Four approved audition takes were reused and 18 new takes completed in the recorded production milestone. Returned voice identities were checked against casting records.

The v3 hero reuses source takes **1, 4, 5, 6, 11, 12, 14, 21 and 22**. It removes the additional policy-question setup and restores the creator's “Different products” question before the guide answers “By starting with real work.” The complete source script and exact timing table accompany this case study.

A continuous WAV carries the dialogue, avoiding separate per-sentence network playback. No time-stretching or pitch shift was used. Captions were first estimated inside takes; that failed. They now use local faster-whisper base.en word timings, reconciled against the approved text, including the product-name tokenization. Full text coverage and non-overlap are checked automatically.

V3 adds an original, very quiet, synthesized stereo ambience bed, created locally from sine tones without sampled music. The voice stays centred. FFmpeg performs mixing and loudness normalization. The target is approximately −16 LUFS with true peak below −1 dBTP; exact encoded-file measurements belong in QC.json rather than being inferred from the WAV.

## Tools and reproducibility

| Activity | Recorded tool or source |
|---|---|
| Brief, script and revision decisions | User/Codex conversation; retained creative treatment and conversational script; external review supplied by the user |
| Animation and editing | React 19.2.3, TypeScript, Remotion 4.0.527 |
| Voice generation | Higgsfield seed_audio; Faye/Ainsley/Isla presets |
| Caption transcription | Local faster-whisper base.en; source-text reconciliation |
| Audio, encoding and inspection | FFmpeg and ffprobe |
| Product evidence inspection | User-authorized Chrome session; no production writes |
| V3 new image/video generation | None |
| V3 sound bed | Locally synthesized original tones; source script included |

The editable project, dependency lockfile, selected audio, logos, captions, scene data and build scripts are included in the source archive. Original raw voice files are retained in the working project. Provider download URLs and private customer data are excluded from the public case-study material.

## Attempts, time and cost

| Item | Supported record | Limit |
|---|---|---|
| Original conversation voice milestone | 18 completed new takes, four reused auditions | Retained-take counts do not inventory all earlier exploratory work |
| Quoted generation cost | 12.6 credits for the 18 new takes | Preflight quote, not a billing receipt; audition costs are outside this figure |
| Hero and v3 paid generation | No new voice/image/video jobs submitted | Does not imply the entire production was free |
| V3 transcription/rendering | Local execution | Electricity, hardware and subscription allocation not recorded |
| Human hours | Not recorded by stage | File timestamps are not labour hours |
| Actual billed total | Not verified | Do not convert the quote to currency or assert a zero total |
| Original voice commercial-use basis | Provider preset voices used under the account | A dated entitlement/licence snapshot was not retained in this package |

The record therefore publishes the known quote and known activity without inventing an actual-cost total. A future production should record receipts and active labour at the time of work. Missing historical measurements cannot be recovered honestly by filling every cell with an estimate.

## QC and release family

V3 includes a clean 1920×1080/30 master, a captioned 1920×1080/30 master, SRT/WebVTT, a rebuilt captioned 1080×1920 version, and a 30-second vertical excerpt. The source package records provenance and checksums. The original 4K portfolio edition remains separate and unchanged.

Verification separates file decoding, dimensions/frame rate, encoded loudness/peak, subtitle coverage, visual state checks and publication. A successful render does not prove creative quality. Automated speech transcription verifies that the Rank attention phrase and the final synthesis are present; it is not a claim of an independent human listening session. No uploaded-platform transcode, public audience screening, legal clearance certification or C2PA signature is claimed.

## Lesson to carry forward

A shorter film is not automatically a clearer film. First align the edit to speech. Then protect a separate area for the product action. Make evidence labels readable, keep illustrative and live demonstrations distinct, and check the actual file being shared. Preserve the failed revisions: they explain the decision better than a claim that the workflow worked perfectly on its first attempt.