Open to GenAI roles in Delhi NCR or remote

Manthan Teotia

Manthan Teotia, Generative AI engineer

I build AI that films ads. As a core engineer of an AI Creative Studio for brands and agencies, I turn one product photo into a 15-second video ad in about 5 to 10 minutes, for about $2. Every video on this page is AI-made: nothing was filmed.

Watch the 90-second filmHow I build AI ads + 4 disasters I fixed
Cost of one 15-second ad
Influencer video$120–300
Creative Studio~$2

Ads made on my pipeline. Tap one to hear it.

Nothing Ear: rush hour, on mute

A call with Maa on a packed Mumbai local, two loud strangers, and one tap that brings her voice back. Hinglish.

Nothing Ear: run the city on mute

A Manhattan jackhammer at 6 am, a tap on the stem, and the run turns into a rock track. 12 s.

PULSR: the flatmate and the dhol

2 AM in a Mumbai PG, a flatmate finds a dhol, and one earbud ends the night's noise. UGC, 16 s.

BREWLO: the 3 AM mystery

A US dorm roommate who never sleeps, and a mini-fridge stacked with cold brew that explains it. 30 s.

KESARI DEW: Mom vs. 12 steps

In Seoul, Umma bins twelve skincare bottles, then steals the one serum her daughter kept. 30 s.

AI presenter reels. One photo of a founder, ten talking-head reels, the same face in every shot.

“I’m not your typical agency”

Her office, one locked look across three clips, captions with the key words in lime, and an end card.

Mohali market: posting isn’t working

A storyboard shot list played out on a market street, a café table and a shop counter.

Five things before a single ad

A natural speaking pace and one gesture per line, landing on its word, across five clips.

Brand campaigns, made on my pipeline. 18 feed ads for five brands, each selling one thing.

Nothing Ear ad: an exploded view of the earbud with labels for the 11 mm driver, ceramic diaphragm, Smart ANC and talk micNothing Ear ad: the earbuds and transparent case on a white studio ground, headline 45 dB of quietNothing Ear ad: a woman at the door of a Mumbai local train, headline Rush hour, on muteNothing Ear ad: a man in the Seoul rain wearing white earbuds, headline Crisper than the rainNothing Ear ad: a runner at dawn, headline 10 minutes. 10 hours.Nothing Ear ad: a city crossing turned to dots around one listener, headline The city, in dots

Ads, generated. Not shot.

A brand used to pay an influencer about $120 to $300 and wait days for one video ad. The Creative Studio makes it in minutes for about $2: roughly 98% less per ad.

Cost per ad
~98% lower
Time to a finished ad
5–10 min
Agencies and brands demoed
12+
Videos used to tune quality
~3,000

How I built it, from input to ad. Six stops, each one shipped to production.

  1. 01

    Input

    A product photo, a brand kit or a winning ad.

  2. 02

    Read

    Gemini Flash reads the reference: 77 s down to 16 s.

  3. 03

    Write

    LLM writers draft three scripts in five beats.

  4. 04

    Film

    Video models film each clip, with a provider fallback.

  5. 05

    Check

    Word checks, drifted clips filmed again, captions on.

  6. 06

    Deliver

    A 15-second ad, ready to run, for about $2.

How the Creative Studio makes an ad

  1. Reads the product or a winning ad

    A product photo or a reference ad goes in. I rebuilt the reference read on Gemini Flash: 77 s down to 16 s.

    A product packshot of graphite earbuds in their case
  2. Writes three scripts in five beats

    From the hook to the call to action, on OpenAI with Gemini as the backup, in English or Hinglish.

    A storyboard of keyframes for an earbuds ad
  3. Films every clip with video models

    Seedance on BytePlus Ark first with KIE as the fallback, Veo and Gemini Omni for Hinglish. The product stays the one in the photo.

    A generated frame of a Mumbai street frozen mid-motion around one woman
  4. Captions, checks and delivers

    51 caption styles, a word-accuracy check and drifted clips filmed again, so the ad arrives ready to run.

    A frame from a finished ad of a mother holding a small serum bottle

The Creative Studio, as a case study

  1. The pain

    D2C brands and agencies paid influencers for every video, waited days, and could not test enough hooks.

  2. What I shipped

    The video engine end to end, from the first port to production and a V2 canvas, built with the BytePlus and Google Gemini teams.

  3. What changed

    Demoed to 12+ agencies and D2C brands across India, the US and Canada, at about $2 an ad.

Forward-deployed · incident log

Fixed by morning. I work where the users are. What broke in production, and how I fixed it.

Read the deep dive: No video may fail Watch the 90-second film
  1. INC-01On callRenders crashing

    Users started uploading long videos

    Broke
    Long source videos pushed the render containers past their memory; the CPU pinned and jobs died mid-render.
    Approach
    Bounded memory and concurrency for every step, a timeout on each step, and a lost job re-submits itself.
    Shipped
    Long videos render. Long sources are now the first test on every render change.
  2. INC-02Vendor escalationVideos stopped

    Our video provider stopped serving videos overnight

    Broke
    The provider we rendered through stopped returning videos, so finished ads never arrived.
    Approach
    Called the BytePlus team that night, walked them through the whole pipeline, then moved our video pipeline to Seedance on Ark directly.
    Shipped
    Ark first, the old provider kept as fallback. Ads kept shipping.
  3. INC-03Field reportAds without captions

    Shoe shelves that looked like captions

    Broke
    A guard meant to stop double captions read a busy shoe-store background as on-screen text, so hooks shipped uncaptioned.
    Approach
    Only text inside the caption region now blocks captions, proved on the customer's own video with a real render.
    Shipped
    Captions decided by where text sits, not by whether any text exists.
  4. INC-04Customer demoSlow first step

    Clone made people wait before anything happened

    Broke
    Reading a reference ad took 77 seconds, and some references could never be cloned at all.
    Approach
    Rebuilt the read on Gemini Flash, and refused an unclonable reference (nobody on camera, no speech) before any money was spent.
    Shipped
    77 s down to 16 s, about 5x faster, and no charge for a reference that can't work.
  5. INC-05Server onlyBroke after deploy

    Worked on my laptop, broke on the server

    Broke
    A hook join was frame-exact locally but gained a frame on the server, which runs a newer ffmpeg.
    Approach
    Every video join now runs in the server's own Linux image before it ships, not just on my machine.
    Shipped
    Joins that match on every machine they run on.
  6. INC-06Pre-launch20 open issues

    20 customer issues, 3 services, 3 days

    Broke
    The pre-launch test round reported 20 issues across the engine, the backend and the frontend.
    Approach
    Test first: a failing test, then the fix, and every change states the fallback that still makes the video.
    Shipped
    All 20 fixed in 3 days and merged after 20 rounds of review.

Features I built into the Studio. Each one started as a client's ask in a demo.

Clone remakes a winning ad with your product

Your product and words inside the reference ad's own timing. Reading the reference fell from 77 s to 16 s.

Hook Variations test five openings

Up to 5 new LLM-written hooks per ad, gated at 75% word accuracy and stitched so no frame repeats.

Hook 1 of 5

51 caption styles that never fail a video

Ranked per video and kept off faces; a caption error falls back and the ad still ships.

A provider outage, solved overnight

When our video provider stopped serving videos, I called the BytePlus team that night and moved the pipeline to Seedance on Ark.

Twelve image models from six providers

Nano Banana, GPT Image, Seedream, Qwen, Grok and FLUX, matched to each creative by ratio, resolution and cost.

Before video: RAG and ML at Eksum

  • Multi-agent RAG over 10,000+ contractsLangGraph agents for answers, summaries and analysis on FAISS search; replies in under 2 s, manual review down 75%.
  • Vendor matching in 2 s, not 30 minEntity extraction, 384-dim FAISS search and scoring: 94% match accuracy, 500+ requests a day.
  • A real-time credit-risk engineXGBoost with SHAP, Isolation Forest fraud checks and Gemini-written credit memos on FastAPI.

From idea to impact. The workflow I build for, and the tools I ship with.

The AI video workflow in six steps: brief, story, visuals, generate with Seedance and video models, refine, deliver, with iterate loops between visuals, generate and refine

Generative AI and LLMs

  • OpenAI
  • Gemini
  • Claude
  • LangChain, LangGraph
  • Hugging Face, LoRA
  • PyTorch

Video and image generation

  • Seedance on BytePlus Ark
  • Veo 3.1, Gemini Omni
  • Nano Banana, GPT Image, Seedream
  • ffmpeg
  • OCR and video QA

Engineering

  • Python
  • FastAPI
  • React
  • PostgreSQL
  • Redis
  • Docker
  • AWS
  • Playwright, pytest