FastAPI service mesh with SDXL Diffusers image generation on AWS A6000 GPU nodes and an NVIDIA Omniverse RTX web viewer, wired into production-facing review workflows. Owned architecture from prototype through production MVP — API design, GPU inference orchestration, asset storage and deployment.
Dushyant
Kashyap
Applied AI engineer with 18+ years building production software for real users, now focused on agentic systems, tool-using workflows and reliable applied AI. I turn brittle, high-pressure workflows into inspectable systems — typed contracts, eval harnesses, audit trails and human-in-the-loop review.
Creative Technical Director and principal engineer with 18+ years across VFX production technology, generative-AI pipelines and global studio workflows. I translate ambiguous creative briefs into production-ready technical plans — leading TDs and engineers, estimating feasibility, and shipping AI-enabled image, video and 3D systems.
Selected work
Shipped → Building → NextAgent-assisted production-intelligence platform for VFX and animation studios. It reasons over bid, timesheet and production-tracking data to detect financial drift, explain operational risk, assemble change-order evidence, run scenario analysis and draft leadership-ready weekly reports — built around inspectable agent workflows with source-linked claims, human approval and an evaluation harness.
India-hosted cloud render farm for VFX, animation and ArchViz studios — a FastAPI + PostgreSQL platform with a provider-agnostic Kubernetes / Argo Workflows render dispatcher, a prepaid credit ledger and a Next.js portal.
Autonomous multi-agent media studio: a durable, model-agnostic orchestration platform where every consequential decision runs through agent committees (propose → critique → verify → adjudicate) under bounded autonomy — typed Pydantic contracts, a constitution-enforced governance layer with spend caps and a tamper-evident audit ledger, and a closed learning loop that promotes proven “plays” cycle over cycle.
LLM-directed creative planning for production: typed planning, multi-agent orchestration, provider routing, image / video generation workflows, live logs and human-in-the-loop review.
A compliance-native election control-room operating system for political campaigns in the US, India and Canada — architected around a policy-engine kernel that enforces multi-jurisdictional election law (DPDPA · ECI · TCPA/FEC · BC PIPA · Quebec Law 25) at the data layer.
Generative content engine that converts a text brief into an engine-ready 3D scene (Unity / glTF), orchestrating LLM + image + image-to-3D models behind typed pipeline contracts with full cost and generation lineage.
Experience
2005 — Present- Architected Wrap.Tech, an agent-assisted production-intelligence platform that detects financial drift, retrieves evidence, runs scenarios and drafts leadership-ready reports with human review.
- Designed Wrap around inspectable agent workflows — source-linked claims, human approval/rejection, a failure-mode taxonomy, deterministic fallback reporting and an evaluation harness.
- Built adjacent agentic prototypes (Backlot and others) exploring typed planning, multi-agent orchestration, provider routing, live logs and fallback paths.
- Shipped the Coca-Cola Content Engine MVP: FastAPI service mesh, SDXL Diffusers on AWS A6000 nodes, an Omniverse RTX web viewer and production-facing review workflows.
- Owned architecture from prototype through production MVP — API design, GPU inference orchestration, asset storage, review integration and reliability trade-offs.
- Tiered 100 TB of footage to cold storage and authored a deployment playbook adopted across global brand work; mentored six TDs on MLOps, USD and AI-assisted workflows.
- Oversaw pipeline delivery for commercials and streaming episodic work; automated QC workflows that saved an estimated 180 artist-hours / week.
- Built and improved internal tools for artists, supervisors and production teams, strengthening hand-offs, review consistency and delivery confidence.
- Led a developer team maintaining the compositing domain across 6 shows; migrated legacy pipeline tools to Ftrack and improved production stability.
- Built studio-wide Nuke QC and delivery tools adopted across production teams; engineered a Photoshop-to-Linux bridge and pipeline suite via a Flask micro-service.
- Integrated USD into the lighting flow and built a suite of reusable PySide tools, cutting shot turnaround by 25%.
- Partnered with lighting, pipeline and show teams to standardize artist-facing tooling and reduce manual scene setup.
Capabilities
Engineering focusCreative-tech focusLLM & Agents
Backend & Infra
Languages
AI / ML
Creative Tech
Production
Education & recognition
Let's build something
worth shipping. Say hello →
Open to senior and principal engineering roles in agentic systems and applied AI. Also available for select pipeline and AI-product consulting through Ayudh.