Voice AI startup · co-founded · 2024 – 2025
Multi-Agent Workflow Automation Platform
Co-Founder & CEO
Multi-agent systemsMARLWorkflow automationAgent coordination
Challenge
Workflow automation with LLMs breaks when one agent must handle every step. Coordination, tool use, and failure recovery need explicit multi-agent design, not ad-hoc prompt chains. The platform had to reflect incentive alignment and robustness under decentralized execution.
Approach
- Engineered an agent-based platform integrating LLM-powered agents for workflow automation with explicit coordination strategies.
- Designed hybrid decision-routing between structured (SQL) and unstructured (RAG) sources so agents pick the right retrieval path per query.
- Partnered with university faculty to benchmark MARL algorithms (MAPPO, HATRPO, HAPPO) for decentralized coordination patterns applicable to production agent networks.
- Proposed and benchmarked SeqPPO (~3× better sampling efficiency vs. MAPPO, HATRPO, and HAPPO), informing scalable multi-agent training and inference design.
Tech stack
Multi-agent orchestration · LangChain · Hybrid RAG + SQL routing · PyTorch · MARL (MAPPO, HATRPO, HAPPO, SeqPPO) · Python · Agent SDK patterns
Outcomes
- Production multi-agent workflow platform, not demo-level LangChain scripts
- Research-backed coordination design (CTDE-style benchmarking)
- SeqPPO contribution to sampling efficiency in multi-agent settings
- Foundation for vertical products (realty, voice) on shared agent runtime
Context
Core platform engineering: shared infrastructure for voice, chat, and vertical workflow products.