Karthik Guduru

Ask me about: holding a 7B model under 400 ms on 8 GB with concurrency · character continuity across AI-generated shots · where approval gates belong in an agent loop (and where they don't)

Building

nights & weekends

v2.9 · invite-only

pod.

The tiny AI film studio. Script in, episode out.

A local app that runs a crew of agents — writer, casting, storyboard, cinematographer, continuity, editor — from script to cut. Shots come from cloud video models; you approve every gate and see the cost before anything renders. Every frame above was made with it.

Watch the showreel ↗ Ask for a key

BrainPuddle

The studio pod. ships from. Also voice agents for Indian businesses, and whatever I'm prototyping this month.

Selected work

tap to play

qARt Codesthe QR is the AR target

Voice AI · in productionVoice agents for BookMyChef

Phone-line agents that answer, qualify and book for a chef & catering marketplace across six Indian cities — telephony + ElevenLabs, built at BrainPuddle. The case study behind "10× faster bookings."

bookmychef.app ↗

GenerativeBreathing life into art

Mona Lisa animated with ComfyUI, a cloned voice, Flux + Kling character motion.

Watch on LinkedIn ↗

Computer visionMatch Your Art

Upload a painting, find who you paint like. FAISS over open-source embeddings.

Source on GitHub ↗

ModelsTrained LoRAs

Two text-to-image LoRAs, open on Hugging Face; a locally trained one powered the TCS icon pipeline.

Hugging Face ↗

MuseumMunch Museum, Oslo

Artwork similarity search across 50 artists and a voice art-mediator for visitors, no AR headset required.

TCS Research, 2023–24

Shipped to production

day job

<400 ms

Private single-GPU LLM stack — Gemma 2B, DeepSeek 7B and Qwen1.5 7B quantized and served with vLLM on one 8 GB GPU. No cloud, for clients who can't send data out.

<50 ms/frame

Real-time CV proctoring — YOLO11n on TorchServe + RF-DETR fine-tuned on 10k+ images. No cloud GPU in the loop.Pearson OnVUE

14 d → 3 d

Autonomous QA agent — regression cycle. AutoGen + Playwright traverses test flows unattended; humans only see the failures.NCR Voyix

2 stacks

Agentic reference stacks on NVIDIA — two production builds on NeMo + Triton Inference Server with CUDA-optimized pipelines, for client deployments.

<1 s

Art similarity search — FAISS across 50 artists, plus a GPT + Whisper voice art-mediator for visitors.Munch Museum, Oslo

Also

Meta XR Hackathon 2021 winner TCS Innovista award · 5G digital twin Luma AI creator program Databricks certified AI agent developer LinkedIn Top Voice · AI

What people say

Working with Karthik was a game changer. He quickly understood our vision and translated it into a sleek AI video, stacking multiple tools to produce an innovative, exciting AI-generated invite. His attention to detail and creative approach truly made our project shine.

Prasanna Krishnan VenkatesaSenior Principal, Research and Engineering · LTM

An exceptional blend of technical expertise, outstanding oratory skills and remarkable leadership. His proficiency in technology is matched by his ability to communicate complex ideas with clarity and confidence.