Aryansh Singh

Clients needed automated workflows on platforms like Spotify for Artists, DoorDash, and Uber Eats. Existing tools:

• Exposed credentials directly to LLMs
• Relied on cloud backends; easily blocked by bot-detection (Cloudflare, etc.)
• Could not orchestrate multiple agents in parallel

Built a privacy-first, multi-agent browser automation platform:

• Parallel multi-agent execution with cron scheduling (daily/weekly/monthly/yearly)
• Fully local backend — credentials never touch LLM context
• CDP-based with fingerprint consistency and behavior simulation for bot-detection resistance
• Headless mode for higher concurrency; live streaming for real-time monitoring
• Multi-LLM support (OpenAI, Anthropic, Gemini, Grok) with structured DOM extraction to cut token costs
• Planner/supervisor LLM validates actions before execution, reducing errors
• Cross-platform: Windows, Apple Silicon, Apple Intel

• Reliable large-scale automation for multiple enterprise clients
• Reduced costs via fully local execution; improved security by isolating credentials
• Higher stability and lower detection rates than previous stacks

Engineering teams spent excessive time on:

• Repetitive Jira → GitHub workflows and manual PR creation
• Navigating repo structure and writing boilerplate for every task
• No traceability between task → branch → PR

Built an AI agent system that:

• Connected via OAuth to GitHub and Jira; read repos, made API calls directly
• Learned repo structure and code patterns automatically
• Created branches, generated convention-aligned code, and raised PRs
• Linked Jira tickets → branches → PRs for full traceability

• Drastically reduced time on repetitive workflow tasks
• Improved code consistency via learned repo-specific patterns
• Bridged product management and engineering with automated traceability

Camera team engineers manually resolved repetitive issues by:

• Searching old tickets for similar bugs and prior CLs (changelists)
• Locating modified files in Perforce — slow turnaround, heavy manual effort

Built an end-to-end RAG-based issue resolution pipeline:

• Embeddings database from historical issues with semantic retrieval
• Partitioned by chipset (LSI/Qualcomm/MediaTek) and device model for precision
• Returned matching issues, CLs, modified files, and resolution patterns
• Integrated with Perforce-based workflows

• Reduced issue-resolution time by ~75%
• Eliminated manual ticket searching; faster onboarding for new engineers
• Created a reusable knowledge retrieval system for the org

DAG-based orchestration (strictly top-to-bottom) broke down for:

• Cyclic agent interactions, iterative refinement, and conditional retries
• Dynamic multi-agent coordination — teams forced to duplicate nodes and build brittle workarounds

Designed a group chat orchestration layer (master-slave) on top of existing DAGs:

• Master agent: tracks goal, progress, and delegates tasks
• Slave agents: handle sub-tasks, dynamically reusable across workflows
• Enabled cyclic loops and node reuse without DAG duplication
• Scoped context per agent — reduced token usage and prevented drift

• Unlocked iterative, cyclic workflows DAGs couldn't express
• Reduced DAG complexity and token costs via scoped context
• Made the platform flexible for real-world enterprise workflows with retries and feedback loops