Tushar Supe

• Built an end-to-end virtual-studio car imaging pipeline: background removal, orientation and tilt correction, AI-generated shadow/reflection, license plate detection and anonymization, and studio-background placement, fusing 15+ production CV models to output photorealistic images at 97–99.5% accuracy or IoU.
• Scaled the imaging pipeline into a high-throughput production system processing ~250K+ images per month with 8× burst capacity, while cutting per-image processing time by 70% and pipeline memory footprint by over 95% through threading, caching, and pipeline optimization.
• Extended the imaging pipeline to bus imagery and shipped a thin-part detection model, raising detection accuracy from 60% to 85% while cutting inference latency by ~4× and container size by ~15× through model and runtime optimization (ONNX export, image slicing, custom post-processing).
• Reduced model training compute cost by 75% through GPU tier optimization and custom training pipeline improvements, and cut distributed-GPU training time from ~3 months to under a week.
• Automated 3D synthetic-data generation (Blender, distributed multi-GPU rendering), cutting a 450+ day rendering workload down to a week and enabling continuous scaling of training data.
• Engineered an interactive 360-degree car spin generation pipeline from walkaround imagery, evaluating and applying multiple novel-view-synthesis and multi-view-geometry approaches spanning classical structure-from-motion, neural rendering, and video stabilization.
• Evaluated and prototyped generative-model approaches for background inpainting, reflection removal, and image harmonization, training custom models for each.

• Drove a company-wide cost-reduction program – in-house model hosting, database and API cost optimization, automated stale-data cleanup, and inference/pipeline efficiency work – delivering over $64,000 per year in combined recurring savings.
• Cut GPU inference hosting cost by ~7× by architecting a self-hosted, multi-model GPU inference platform across two cross-cloud migrations, reducing reliance on paid managed ML hosting.
• Led a two-phase authentication performance overhaul on a core production API, cutting response time by over 99% to under 100ms through database indexing, caching, and query redesign.
• Implemented a GCP AI inference gateway across 30+ in-house inference endpoints with a Redis-based rate limiter and custom metrics-driven autoscaling, handling ~1,000 requests per 20-second window while optimizing for cost-efficiency over raw throughput to match real production load and minimize idle spend.
• Established MongoDB schemas, aggregation pipelines, and indexing strategy for a document database layer supporting multiple microservices, and built the asynchronous task queue infrastructure behind it, resolving recurring worker crashes, memory exhaustion, and message-duplication issues under production load.
• Owned end-to-end production incident response and database reliability at scale – including a multi-month retention cleanup that reduced one major collection's size by a third – across application, database, task queue, and third-party API layers as the platform's primary infrastructure engineer.

• Advanced from individual contributor to a technical lead role at a sub-10-engineer startup: mentored four AI and backend engineers – while still independently shipping individual technical work end-to-end.
• Owned the AI and computer vision technical roadmap and technical hiring for AI engineering roles; handled technical support escalations for major enterprise accounts.
• Guided a team through delivery of an automated social media content pipeline (AI avatar-video, script, and voice generation with multi-platform posting and comment-reply agents) and shipped internal analytics dashboards covering image quality, volume, failure rate, and team performance.
• Currently steering a team building an automated image quality-check and auto-correction system that flags issues via AI and routes unresolved cases to human review.

Mechanical plant maintenance at a large-scale manufacturing facility.

Agentic system automating the end-to-end job application workflow: searches jobs, parses job descriptions, matches them against a structured candidate profile, generates tailored application materials, and completes submission steps through browser automation. Building with LangGraph for agent orchestration, retrieval-augmented generation over candidate and job data, tool-calling and MCP integration, and a FastAPI/PostgreSQL backend with Playwright for browser automation.