Vishnu Anilkumar

Adroom Creative Playground — multi-modal GenAI platform

• Built AdRoom, an AI-powered creative automation platform from inception to $2M+ ARR; led end-to-end product development including roadmap, design, sprints, and a 4-person team through v1 launch.
• Architected and operate a 46-route creative-generation platform across 5 independently deployed services, orchestrating 25+ Baseten-hosted models (SDXL, ControlNet, custom Flux fine-tunes, LaMa inpainting, harmonization, upscaling) plus FAL, Replicate, OpenAI, Anthropic, and Google Gemini APIs behind a unified provider abstraction with webhook-driven async fan-out.
• Designed the unified_generation orchestrator and its four execution variants — unified (single brief), controlled (knob-per-stage power mode), template-based (brand-sealed templates), and bulk_execution_controller (1,000+ variants fanned out per call with parent-child task aggregation, per-task retry, and NSFW gating).
• Built the Ad Copy Generation engine (llm_advertisements) — brand-voice-grounded copy across FB / IG / TikTok / Pinterest / LinkedIn shapes with per-provider routing (GPT-4o for CTAs, Claude for narrative, Gemini for batch) and draft/polish cost-tiering — 60%+ cost reduction vs all-flagship with no quality regression.
• Shipped the Prompt Intelligence layer (prompt_enhancer + prompt_variation) — parses a 5-word user brief into a 5-field ad concept, fans out 12 variants with embedding-based semantic diversity scoring, and compiles per-stage prompts for SDXL / Flux / DALL-E / video pipelines; outputs scored back via the creative-QA MCP for closed-loop quality.
• Owned the AI Image Editor suite — 7 operations (image_editor, lama_cleaner, generative_fill, enhance_by_reference, harmonizer, segment_anything, blend) with SAM-driven auto-masking so users no longer hand-paint masks; harmonizer matches color temperature and light direction across a campaign series for visual cohesion; mask → fill → harmonize → upscale chainable under a single async task.
• Built the Product-to-Video studio — 4-stage pipeline (image_to_video → product_to_video → camera_movement → extend_video) with vision-LLM continuity checks between stages and automatic Baseten → FAL → Replicate provider failover on cold starts.
• Built the UGC video ad pipeline — talking-human avatars + ElevenLabs voice cloning muxed via ffmpeg with lip-sync correction; Celery + RabbitMQ queue scaled to thousands of videos per execution batch with per-task retry.
• Drove production LLM cost & latency optimization through context management, prompt caching, per-tenant token budgets, and New Relic-backed observability across the multi-provider orchestration layer. Stack: FastAPI, ARQ, Redis, RabbitMQ, PostgreSQL (asyncpg + SQLAlchemy 2), AWS EKS + S3 + CloudFront, New Relic APM.

Competitor Insights — multi-tenant LLM analytics service with MCP exposure

• Designed and shipped a workspace-scoped microservice that ingests Meta Ad Library snapshots and produces 20+ LLM-driven strategy reports per snapshot — using GPT-4o-mini for text reports, GPT-4o vision for opening-frame and creative analysis, OpenAI o-series reasoning models for executive summaries and key takeaways, and Whisper-1 for video-ad audio transcription.
• Built an MCP server exposing 11 read-only tools over Streamable HTTP, integrated with Claude Desktop and Cursor — extends the same data backend to AI copilots without duplicating business logic, with API-key auth, context resolution, and SQL-injection guards. Shipped a second Creative QA MCP (8 tools: brand-compliance, heatmap analysis, cross-creative pattern detection) with multi-modal routing across Claude Vision, GPT-4o, and Gemini.
• Implemented credit-aware execution (per-report pricing with media-type weighting), recurring sync and report schedules (cron-driven, every 6h), self-healing for stuck jobs with Slack alerting, 9 zero-downtime SQL migrations (report dedup, schedule uniqueness, job-run enums, full table restructure), and Gamma.app-powered PDF / PPTX export for client deliverables. Stack: Python, FastAPI, FastMCP 3.1, ARQ, PostgreSQL, Redis, RabbitMQ (Kombu fanout), S3 + CloudFront, ffmpeg.

Cross-platform engineering leadership

• Designed and deployed a RAG-based LLM agent for pandas data operations — owned dataset design, training, evaluation, and serving — automating analytical workflows for non-technical users.
• Define the ML technical roadmap and lead cross-team design review of models, datasets, and evaluation patterns; drive MLOps discipline across both products — versioning, drift monitoring, GPU autoscaling, release gates, and rollback strategy on production inference endpoints.

• Dawn Foods (Customer Lifetime Value): Led the engagement end-to-end as Senior MLE; delivered a production CLTV model with monthly revenue as the North Star metric — covered data onboarding, modeling, and stakeholder sign-off.
• Dyson UK (Production ML pipeline): Designed and deployed a Kubeflow pipeline on GCP AI Platform with custom model monitoring, delivering an ensemble-based CLTV model into production.
• Mentored a team of seven MLEs across multiple engagements; advised R&D and product teams in marketing analytics. Recognised as Employee of the Year for delivery quality and team leadership.

• Skill-extraction CNN: Developed and deployed an entity-recognition model for an enterprise hiring platform — reached 82% accuracy and materially improved skill detection and candidate matching.
• Legal-clause semantic search: Built a BERT + Elasticsearch search engine enabling faster clause comparison and retrieval across large legal corpora.
• Signature and stamp-seal detection (YOLOv3): Computer-vision model automating proofreading of legal documents.

• Resume intelligence platform (client-facing): Candidate scoring, section identification, content-quality evaluation, and alternative-industry / role prediction using TensorFlow and Random Forest; fine-tuned BERT for role-prediction from experience profiles.
• Customer-churn prediction: Achieved 82% accuracy across 22,000 customer records for a financial-services client using XGBoost; supported targeted retention campaigns.
• Cattle face-identification: Computer-vision model automating cattle insurance claim verification for India's largest livestock insurer.

• Credit-risk analytics: Probability-of-default modelling from KYC data using Regression, SVM, Random Forest, and XGBoost with SMOTE, anomaly detection, and feature-selection pipelines.