<p>Owned mission-critical GPU-execution infrastructure and contributed across generative pipelines, computer vision, automated QA, prompt engineering, and MLOps.</p><p> Designed and built from scratch a multi-provider GPU workflow-execution orchestrator (TypeScript/NestJS): one queued, fault-tolerant REST API abstracting a direct ComfyUI instance and two serverless GPU clouds behind a single job model.</p><p> Implemented fault tolerance: FIFO dispatch state machine, cooperative cancellation across three provider models, automatic cross-provider failover with restart-safe retry lineage, and SQLite-backed crash recovery/reconciliation on boot.</p><p> Built graph-theoretic live-progress modeling — critical-path computation (topological sort + longest-path DP) plus learned per-node timing profiles — for smooth, monotonic, ETA-bearing progress streamed over WebSocket.</p><p> Delivered the contracted automated QA system for 3D & Gen-AI outputs: a self-contained GPU service aggregating logo/brand-integrity, on-screen-text OCR, identity, sharpness/compression, face-consistency and vision-language checks into one score/verdict/report.</p><p> Authored a research-grade video & audio QC toolkit (sharpness, temporal flicker, IQA, compression artifacts, face consistency, mangled-text gate; loudness/LUFS, clipping, SNR, intelligibility) with per-method licensing/risk analysis and deterministic self-tests.</p><p> Engineered a controllable diffusion background/relighting pipeline (monocular-geometry mesh reconstruction → LoRA artifact repair → physically-controllable relighting) and a distributed, SSH-less Blender render farm generating harmonization training data with cost-aware GPU autoscaling and guaranteed teardown.</p><p> Built an interpretable two-phase camera-pose solver (Umeyama similarity alignment + reprojection refinement) for 3D-consistent compositing, with dual TypeScript/Python implementations kept in lockstep via golden-fixture parity tests.</p><p> Created an ontology-driven, schema-validated cinematography prompt engine and a typed DAG node framework shared between a Python back-end and a React node-graph UI; added LLM/VLM-assisted user tooling.</p><p> Established MLOps foundations: reproducible checksum-pinned baked GPU Docker images, provider-abstracted serverless deployment (RunPod/Vast/serverless-container), and automated cloud-GPU price/performance benchmarking with hard cost guards.</p><p> Integrated the orchestrator on both the Python worker side and the AdonisJS 6 + PostgreSQL platform back-end; worked within an enterprise multi-tenant SaaS (ACL, credits, campaigns, Auth0, GCS/R2, React 19 + Fabric.js front-end).</p>
