Aurum Quanta Chat.
Assistants and routing systems that hand off to a human when they should.
Assistants, chatbots, and routing systems for customer support, internal queries, and operational triage. The hard part is the handoff: when the bot's confidence drops, the conversation moves to a human with the full history attached.
Multi-turn context, grounded retrieval, intent detection, and sentiment-aware routing. Wired into your CRM, helpdesk, or comms stack on the first sprint, because the integration is usually where these projects fail.
One query in. Watch the agent reason, call tools, and respond.
Click Run trace below to see how the agent decomposes the query, calls tools, and composes the answer.
Pick a sample query. The trace animates the actual ReAct loop a production agent runs: parse the question, decide which tools to call, observe the results, and compose the answer. The conversation panel shows what the user sees; the tool trace shows what actually happened underneath.
Note · this is a simplified demo
The traces here are deterministic for a fixed set of sample queries. A real engagement would wire the same loop to actual tools (search, retrieval, internal APIs, function calls into your business logic), with retry policies, fallback behaviour, output validation, prompt-injection guards, conversation memory, and per-tool eval harnesses. The shape of the loop above is what production looks like; the depth comes from doing it on your tools, your data, your conversation policies.
How an agent actually reasons.
Real agents don't answer in one shot. They alternate between thinking, calling tools, observing the result, and thinking again - until they have enough to respond.
Concrete deliverables.
Customer support bots
Multi-turn context, grounded retrieval, and a clean handoff to a human when the bot is out of its depth. The full conversation history goes with the handoff so the customer doesn't repeat themselves.
Internal knowledge assistants
Connects to your wiki, CRM, and internal docs. Permissions are respected at retrieval time, so what a user can see in the assistant matches what they can see in the source system.
Intent and sentiment-aware routing
A frustrated customer or a high-stakes query (complaint, legal, billing dispute) escalates immediately. Everything else routes to the cheapest path that handles it well.
CRM and helpdesk integration
Zendesk, Salesforce, HubSpot, Intercom, Slack, Teams, or whatever stack you've already built. We don't ask you to migrate.
Routing isn't one threshold. It's a stack of signals.
# chat/route.py: intent · sentiment · history · tool budget
def route(turn: Turn, history: list[Turn], session: Session) -> Reply:
intent, conf = classifier.predict(turn.text, history=history[-3:])
sentiment = sentiment_score(turn.text, history=history[-3:])
# Non-negotiable: certain intents never get a bot response
if intent in {"complaint", "legal", "billing_dispute"}:
return escalate(turn, reason=f"intent={intent}")
# A frustrated user overrides every other signal
if sentiment < -0.5:
return escalate(turn, reason=f"sentiment={sentiment:.2f}")
# Low intent confidence: the bot doesn't know what it's looking at
if conf < 0.7:
return escalate(turn, reason=f"low_conf={conf:.2f}")
# Stuck-in-a-loop guard: tool budget per session
if session.tools_used >= MAX_TOOLS_PER_SESSION:
return escalate(turn, reason="tool_budget_exhausted")
return handlers[intent](turn, history=history, session=session)Four independent signals can each trigger escalation. The bot answers only when all four agree it should.
How it would unfold.
Design
Conversation design, integration audit, success metrics (deflection, CSAT, escalation) agreed.
Pilot
Single-journey prototype behind a feature flag, limited user base, measured against baseline.
Rollout
Full user base, fallback strategies, escalation workflows, monitoring dashboards.
Ongoing
Conversation quality review, intent tuning, and expansion to new journeys.
Tools we reach for on this kind of work.
Common questions.
What happens when the bot can't answer?
It escalates with the full conversation history attached. The customer doesn't have to repeat themselves and the agent picks up exactly where the bot left off.
Can it speak languages other than English?
Yes, though quality varies by language and model. We test on your users' real messages. Synthetic benchmarks tend to flatter the model and miss the failures you'd actually see in production. If a particular language is going to need more work, we'll say so upfront.
How do we know if it's actually helping?
We measure deflection rate, customer satisfaction, and escalation rate against the human-only baseline. The numbers and the windows are agreed in writing before any build starts.
Let's build it.
A 30-minute discovery call. We'll tell you whether we're the right shop for this.
Book a discovery call →