chandima
← all projects / 01

chatstream

AI agents that answer customer messages on WhatsApp, Instagram, Messenger and the web, and hand over to a person when they should.

[ open chatstream.us ↗ ]
The ChatStream home page: Automate Customer Conversations Across Every Channel above a row of messaging platforms.
role
solution architecture · full-stack
scope
omnichannel ai platform
year
2026

typescript · durable queues · llm routing · meta webhooks

telemetry // numbers sample telemetry · replace with live feed
messages / mo
420k+
4 unified channels
p50 route latency
148ms
edge webhook → queue
auto-resolved
82.4%
clean human escalation
webhook delivery
99.98%
idempotent zero-drop

routing & handover latency

24h burst profile · p50 edge ingest vs p95 agent reply

p50 ingest p95 llm + tools
600ms 300ms 0ms 00h 04h 08h 12h 16h 20h 24h p50 148ms

channel volume distribution

  • whatsapp 54%
  • instagram 26%
  • web widget 13%
  • messenger 7%

omnichannel ingestion & handoff topology

idempotent envelope → durable queue → guarded agent → human desk

hooks key ack auto >= 0.85 escalate [01] channels wa · ig · web · fb [02] edge ingest verify + dedupe [03] durable queue ordered per thread [04] agent router rag + tool policy [05a] llm inference + kb grounded reply stream [05b] live human inbox state summary handoff

01 / context

Support and sales teams juggle incoming threads across WhatsApp, Instagram DMs, Facebook Messenger, and web chat. Most automation tools either reply with rigid decision trees or let an LLM improvise past the point where a human should step in.

ChatStream unifies every channel into one conversational pipeline where agents resolve routine inquiries quickly and escalate cleanly with full context.

02 / architecture

Inbound platform webhooks normalize into a single idempotent message envelope before entering durable queues. Retries, burst traffic, and duplicate Meta deliveries never create double replies or dropped customer threads.

Each workspace configures knowledge sources, tool calls, and explicit handoff boundaries so the moment confidence drops or a customer asks for a person, the thread routes straight to a live inbox.

  • unified ingestion across whatsapp, instagram, messenger, and web
  • idempotent webhook queues that never drop or duplicate turns
  • explicit human handoff with summarized thread state

03 / craft

The hardest part of an AI messaging product is trust: operators need to see why an agent answered a question, override it cleanly, and test changes before exposing customers to them.

The interface keeps channel noise out of the way so a human operator can step into an active thread mid-conversation without repeating questions the customer already answered.

04 / outcome

Live at chatstream.us, powering multi-channel customer conversations with predictable escalation paths and resilient webhook delivery.