Hours
Weeks
Commits
Role
Resolo
Multi‑Agent AI Support Orchestration With Human Escalation With a Twist
Resolo is a multi tenant AI customer support and sales platform that routes conversations across website widgets, social channels, and WhatsApp while keeping humans in control when it matters.
resolo-ai/resolo
Resolo is a multi tenant AI customer support and sales platform that routes conversations across website widgets, social channels, and WhatsApp while keeping humans in control when it matters. It exists for teams that outgrew simple chatbots and need AI agents that can answer from organization specific knowledge, collaborate with human agents, and respect real world constraints like agent workload and message rate limits. Resolo replaces ad hoc scripts and disconnected tools with a single system that can answer FAQs, run outreach campaigns, and escalate tricky cases to humans without losing context. Under the hood, Resolo uses a custom multi agent orchestration engine in TypeScript with Next.js and React that tags model responses with confidence scores, explicit handoff markers, and escalation signals so conversations can move safely between AI specialists and humans. An organization scoped RAG layer stores embeddings in Postgres pgvector with a Redis cache, which keeps semantic search close to transactional data and avoids a separate vector database. Redis also coordinates round robin escalation, takeover events, and notifications, while an API keyed chat widget validates domains and streams Vercel AI SDK responses to third party sites. Billing is treated as a first class concern: AI token usage is converted into credits per model, mapped to plans, and synced to Polar so feature access, campaigns, and channels stay aligned with cost. In production, Resolo has handled over 100k interactions across web and social channels, maintains predictable agent workload through Redis backed rotation, and delivers AI responses that users rate at 4.9 on average, which demonstrates both technical soundness and real customer value.
Designed a multi-agent orchestration engine using confidence tags, explicit [HANDOFF:*] markers, and escalation signals to route conversations between AI specialists and humans with controlled fallbacks.
["Designed a multi-agent orchestration engine using confidence tags, explicit [HANDOFF:*] markers, and escalation signals to route conversations between AI specialists and humans with controlled fallbacks.","Built an organization-scoped RAG layer on Postgres pgvector with Redis embedding cache - enabling fast, relevant retrieval across FAQs and instructions without introducing a separate vector database.","Engineered an auto-evolving FAQ pipeline that classifies guidance, performs semantic deduplication via embeddings, and updates organization FAQs based on configurable confidence thresholds to continuously improve support knowledge.","Implemented a Redis-backed round-robin escalation system that assigns conversations to eligible agents, persists rotation state, and publishes per-user notifications for predictable and transparent workload distribution.","Developed a widget-focused public chat API that validates per-organization API keys, enforces domain whitelisting, and manages conversation lifecycles while streaming AI responses securely to third-party sites.","Created an AI-aware conversation takeover flow that lets humans temporarily control threads, then safely re-hand off to AI agents with background responses triggered when user messages are pending.","Built a WhatsApp campaign execution engine that batches sends, respects Meta rate limits, retries failures with backoff, and records detailed send states for reliable large-scale outreach.","Designed a credit-based billing system that converts per-model token usage into plan-specific credits, tracks overages, and syncs usage events with Polar to align AI costs with revenue."]
Engineered an auto-evolving FAQ pipeline that classifies guidance, performs semantic deduplication via embeddings, and updates organization FAQs based on configurable confidence thresholds to continuously improve support knowledge.
Implemented a Redis-backed round-robin escalation system that assigns conversations to eligible agents, persists rotation state, and publishes per-user notifications for predictable and transparent workload distribution.
Developed a widget-focused public chat API that validates per-organization API keys, enforces domain whitelisting, and manages conversation lifecycles while streaming AI responses securely to third-party sites.
Created an AI-aware conversation takeover flow that lets humans temporarily control threads, then safely re-hand off to AI agents with background responses triggered when user messages are pending.
Built a WhatsApp campaign execution engine that batches sends, respects Meta rate limits, retries failures with backoff, and records detailed send states for reliable large-scale outreach.
Designed a credit-based billing system that converts per-model token usage into plan-specific credits, tracks overages, and syncs usage events with Polar to align AI costs with revenue.
Resolo gives support and marketing teams an AI-first, multi-channel inbox that automates customer conversations while preserving human control, reliable delivery, and accurate usage-based billing.