Brief
Where Dvar fits in Kuku's AI storytelling stack
Five needs from what Kuku has said in public, and the Dvar piece for each. Dvar is the measurement layer that makes an in-house AI team faster and its spend accountable: it does not write the stories.
| What Kuku has said | What that needs | Dvar |
|---|---|---|
Generative AI at every layer: adapting stories to the Indian context, scripts, dialogue, thumbnails and descriptions, with custom GPTs for each process. | Every AI-written scene, dub and thumbnail line needs standards checks, at a volume no review team can read. | Content checks: free code checks, a small open model on Dvar for every scene, frontier only when the small model is unsure. Each verdict points at the line. Content checks → |
In-house recommendation system since 2021–22 (homepage and More Like This), built after rejecting a vendor; Databricks on Google Cloud with MLflow for the recommendation models. | Recommendations and search are built in-house. Hinglish, misspellings, Devanagari and described asks are normal queries, and every model change needs a number before it ships. | Search evals: graded queries by type (title, misspelling, Devanagari, described), nDCG, recall and zero-result rate, plus cost and p95 latency for any model step. Search evals → |
English writing moved fully to AI; Hindi and Tamil quality described as decent; Telugu, Kannada, Malayalam, Bangla and Marathi called still poor (dated). | Quality differs by language and genre. The question for every draft model is quality per language and cost per accepted draft, re-run as models change. | Generator evals and a release gate: every prompt or model change read against briefs; language, rating, cast and residue are critical; a judge reads character and cliffhanger. Generation evals → |
KukuTV: 200–250 shows a month, targeting 1,000. AI for story-potential scoring, script fixing, pre-visualisation, hook and cliffhanger identification, VFX, dubbing and music; dramatic language in Indian languages still not good enough for end-to-end dubbing. | Dubbing into Indian languages is where quality is weakest and volume is growing. Each language needs its own number, not one average. | Localisation evals: script, names, Hindi or English left in and timing in code; meaning and tone by a judge; cost per accepted dub per language. Localisation evals → |
Genre-specific model routing, self-reported by the CEO: certain models handle science fiction, others true crime. How that ranking is measured is not public. | Models are already routed by genre. A gateway makes that routing explicit, priced per call, and decided by evals rather than by feel. | One gateway: each workflow on the cheapest model that clears its eval, with escalation and fallback, every call traced back into the evals. Open models served from India, priced in rupees. Gateway → |
Questions for Kuku
The meeting agenda: what we need to know to turn this demo into a pilot on your data.
- 1ChecksHow many scenes, episodes and dubs a month go through standards review today, and which failures reach listeners most often?
- 2ChecksWho sets the bar per age rating, and would your standards team label 300 scenes as the first gold set?
- 3SearchHow is search quality judged today: offline sets, click-through, or both? Which query types fail most: misspellings, Devanagari, or described asks?
- 4GenerationWhich languages and genres matter most for the script generator in the next two quarters, and how is a prompt or model change approved today?
- 5LocalisationWhich language pairs are in production for dubs, and how is timing matched to the original today?
- 6CostRoughly how is model spend split across drafting, checks, search and dubs, and which provider bills are growing fastest?
- 7OwnershipWho owns evals today, and would a post-trained model Kuku owns for one workflow be worth a pilot?
- 8ProbeWhere does TTS sit in the pipeline, and is voice quality measured per language? (A probe for later; not in this demo.)
A four-week pilot, starting with content checks
Week 1
Connect
One line to route the checks workflow through Dvar's gateway, or traces from your own logs. 300 scenes labelled with your standards team: the gold set.
Week 2
Measure
Every model on the gold set: catch rate per check, false alarms, cost per 1,000 scenes and latency. Your current setup is the baseline.
Week 3
Route
The tiered route live on one show line, beside today's process. Search and generator evals set up on the same pattern.
Week 4
Decide
Cost and quality per workflow, on your data. Where a gap repeats, a post-trained model Kuku owns, released only through the gate.
What stays with Kuku
Your standards team sets the bar and signs off the gold set. Scripts and traces stay in your environment or on Dvar servers in India. Models trained on your data are yours. Creative calls stay with writers: the checks flag a line, they do not rewrite it.