Software first. Generative models only become usable in production because of the agents, services and interfaces around them, and that is the part I write.
Conductor/orchestrator agents that resolve project context and delegate to specialist skills with custom tools.
FastAPI services with async jobs, SSE progress, job-router integration and state machines hardened against retries and stale workers.
React interfaces for AI Studio Labs tools and Projects V2, with sessions, history, deep links and per-user state that follows you across devices.
LoRA fine-tuning, ComfyUI pipelines and nine image/video models wired into production, from FLUX and Seedream to Kling, Veo and WAN.
Software I have designed and shipped, the generative pipelines it runs for film, and earlier ML research.
A multi-agent film-production system I designed and built end to end. A FastAPI backend runs a conductor agent that resolves project context and delegates to specialist agents with their own tools and skills; a React front end streams their progress over SSE and lets the team edit storyboards before any expensive generation runs.
AI Studio Labs is Galleri5's shelf of production tools for the film team. I contributed backend (Python, FastAPI) and frontend (React) work across them, including Texture Studio, Asset Kit, Light Card, Camera Forensics, Wall Studio, Style Studio and Motion Previs.
AI Studio Labs · production tools
Projects V2 is Galleri5's production tracker for AI feature films, keeping every scene, shot, task and asset of a film in one place. I maintained it and helped the team run production on it, and added the Activity Feed tab, task links between Labs tools and production tasks, and render-cost estimates so budgets reflect real spend.
Every delivered shot of an AI feature film had to be traced back to the AI generation that produced it. I built the matcher: an MD5 pass catches exact copies, then each clip is embedded whole with gemini-embedding-2 and searched by cosine similarity in Azure AI Search, with a side-by-side review viewer for the uncertain ones.
An LLM agent that drafts on-brand replies to Telegram support questions, routes each draft through human review, and posts approved answers. FastAPI and PostgreSQL sit behind a Streamlit dashboard that tracks response times.
Review dashboard · human in the loop
An AI research assistant that ingests large document collections, extracts entities and relationships with LLMs, and turns them into interactive knowledge graphs, so researchers can see how concepts connect across papers.
Walkthrough · 3:55A personalised Indian meal planner built on a PostgreSQL database of 5,000+ recipes with cleaned text and standardised ingredients. A short profile quiz feeds a classifier, RAG prompting and GPT-2 / LLaMA 3 fine-tuning experiments.
Built the character and style generation pipelines for an AI feature film, and trained LoRA adapters on GPU servers with AI Toolkit and Hugging Face, so the same face, costume and look hold across every asset.
Tools the production team uses daily. GPT-powered storyboard generation turns a shot list into framed boards; background and foreground replacement puts characters into new plates; style transfer brings every frame to the film's look.
Reusable pipelines for character consistency, style control, image-to-image, image-to-video and video generation across FLUX, Seedream, Seedance, WAN, VACE, LTX, Veo, Kling and Z-Image, tuned and supported across productions.
ComfyUI · Qwen-Image + style LoRA
A reference image in, a usable 3D asset out. Trellis and Tripo build the mesh from concept art, Nano Banana cleans up the source views, and Blender handles retopology, texturing and export.
Two models built on a Swin Transformer backbone. The detector feeds four Swin stages into a feature pyramid network and a ResNet detection head, reaching 96% accuracy, a 40% improvement over existing baselines. The second is a Swin encoder–decoder with skip connections for pixel-level segmentation at 512×512. Full pipeline: OpenCV preprocessing, augmentation, transfer learning and mixed-precision training.
Detection architecture · Swin + FPN + ResNet head
Integrating Artificial Intelligence Aided Vibrational Spectroscopy Applied to Medical Biopolymer Applications. In Vibrational Spectroscopy for Biopolymer Characterization, pp. 237–279, Springer, Cham, 2026. Third author of six.
Two kinds of LoRA I train: character LoRAs that keep a face and costume identical from shot to shot, and style LoRAs that give a film its look. The frames below are style LoRA examples.
The characters belong to productions that are not yet released, so this work is described here rather than shown.
From ML research internships to shipping AI software for film production.
Published work and formal AI training.
Panicker A., Manoharan K., Thambi S., Gangadharan S. V., Poornachandran P., Nambudiri M. K. In Vibrational Spectroscopy for Biopolymer Characterization, pp. 237–279. Springer, Cham, 2026.
Read on SpringerLink ↗Certificate of completion from 100xEngineers' applied generative AI programme, instructed by Siddhant Goswami and Sridev Ramesh.
AI software engineering, from the model layer to the interface.
I write the software that makes generative AI usable in production: the agents, the APIs, the job pipelines and the screens a crew works in.
At Galleri5 I work across the stack: FastAPI services and async jobs on the backend, React interfaces on the front, and the model layer underneath, from LoRA training to ComfyUI pipelines across nine image and video models.
My largest build is Hermes, a multi-agent system that reads a screenplay and plans the shots. Before Galleri5 I built computer-vision and time-series models in research internships, and co-authored a Springer chapter on applied AI.