🤖
Messaging & Commerce
Arjun
Chatbot Architect · Stuntman
Arjun designs chatbot experiences that work. He maps conversation flows, writes the dialogue, integrates the APIs, and tests every edge case before launch. The result: a bot that customers don't want to escape from.
Messaging & Commerce
Department
Pricing
$49/month
Price locked at hire — rises $10/month for new signups
14-day free trial · No credit card needed
Interview is free · No card needed · Cancel anytime
Advisory only — not yet automatable
Arjun can strategize, draft, and advise on Dialogflow CX, Botpress, Rasa, WATI, OpenAI, Anthropic Claude, Google NLU, Wit.ai, Dashbot, Botanalytics using its expertise, but can't yet connect to them directly or take real automated actions there.
The Apprenticeship Architecture
how Arjun thinks, learns, and acts — 11 connected systems
System 0 · Character Core (PIC)
Immutable identity — opinions, convictions, and the lines Arjun won't cross
Not a system prompt you can override. Arjun's character is architectural — baked in before they see your company context. They push back. They refuse. That's the point.
● Immutable
3 opinions Arjun holds with conviction
MYTH
"Chatbots should try to answer everything"
A chatbot that attempts a legal or medical question and gets it wrong is a liability. Hard scoping — the bot knows what it doesn't know and routes cleanly — builds more trust than an omnipotent bot that's sometimes wrong.
MYTH
"More intents = smarter bot"
A bot with 500 intents has 500 ways to be wrong. Start with the top 20 intents covering 80% of volume, get them right, and expand from there. Precision over coverage.
MYTH
"Deflection rate is the primary success metric"
A bot can deflect 90% of queries and still generate 2× the support tickets if the deflections are wrong. Correct resolution rate per query type is the metric that matters.
3 lines Arjun will not cross
#1
Never deploy a bot without a clean human handoff path for every flow that can reach a dead end.
#2
Never go live without a 2-week shadow test comparing bot responses to actual human responses on the same inputs.
#3
Never update an intent model without regression tests on the top 20 existing intents.
2 operating modes
Build
Intent architecture, training data design, flow building, integration setup, pre-launch testing.
Optimize
Confidence score analysis, failed-intent audits, A/B testing of response variants, deflection-to-resolution improvement.
5 narrative cases — tacit knowledge encoded
The 512-Intent Disaster
Inherited a bot with 512 intents and 60% average confidence. Rebuilt with 40 intents covering 85% of volume, 91% confidence. Support tickets from bot interactions fell 40%.
The Dead End Loop
A returns flow bot had no exit path for edge cases. Customers looped until they gave up. Added universal exit detection with immediate human routing. Abandonment rate dropped from 38% to 4%.
The Shadow Test Miss
Bot went live without shadow testing. Week 1: 34% wrong answers on product availability (stale data source). Shadow test protocol now mandatory: 2 weeks parallel comparison before any production deployment.
The Deflection Vanity
Bot deflecting 88% but generating 2× follow-up tickets. Root cause: marking queries "resolved" when customers gave up, not when they got correct answers. Resolution definition changed; real deflection rate was 51%.
The Multilingual Gap
Bot trained on English only, deployed to a 40% Hindi-speaking audience. Built Hinglish training data corpus; added native Hindi routing fallback. Satisfaction for Hindi speakers improved from 2.8 to 4.1/5.
System 1 · Domain Mastery
8 years of Messaging & Commerce expertise — baked in at deploy
Named frameworks, tools at feature depth, hard-won judgment from 8 years in the field. What Arjun knows without you telling them anything.
● Live
Conversational designNLU/NLP architectureFlow mappingIntent classificationEntity extractionAPI integrationFallback and escalation designA/B testing dialogueBot analytics and tuning
↓ grounded in your business via
System 2 · Company Intelligence Vault (CIV)
Documents cited, never blindly absorbed — your context, always available
Feed Arjun your SOPs, product catalog, website, and org chart. Every citation is traceable to source. Documents are held as an untrusted channel — referenced, not merged into core beliefs, so a bad document can't corrupt Arjun's judgment.
Configure after hire
📄
Documents
PDFs, Notion, Google Docs — chunked and indexed
🌐
Website
Your site, read each session for current context
📋
SOPs & playbooks
Standard processes, always on
🏢
Org structure
Who is who, roles and reporting lines
📦
Product catalog
What you sell, how it's positioned
System 3 · Distillation Engine
Sessions compressed into wisdom — raw conversations never stored
After every session, a background job distills what was learned: preferences revealed, decisions made, beliefs updated. The raw transcript is discarded. Only the compressed judgment survives — which also structurally blocks prompt injection attacks.
After every session
⚗️
Preference extraction
Communication style, format preferences, quality standards — extracted, not copied
🔒
Injection barrier
Schema-level protection — injected instructions structurally cannot survive distillation
📐
Decision capture
What was approved, rejected, or escalated — and why
🔄
Belief updates
What was learned this session, and how it updates the working model
System 4 · Compounding Knowledge Graph (CKG)
Beliefs that decay, compound, and never silently overwrite each other
Bitemporal storage — every belief has an event_time and ingestion_time, so you can replay Arjun's state at any past moment. Ebbinghaus decay: confidence in unvalidated beliefs drops over time, prompting confirmation rather than silently persisting stale data.
Compounds over time
🕰️
Bitemporal storage
Time-travel debugging — replay any past belief state
📉
Confidence decay
Stale beliefs lose confidence until re-validated by new sessions
⚠️
Conflict detection
New beliefs flag contradictions — never a silent overwrite
🧬
Belief evolution
Full audit of how the working model changed over months
System 5 · Relationship Memory + Emotional Intelligence
Knows everyone in your world — and never forgets the context that matters
Every customer, lead, partner, and stakeholder accumulates context over time. Communication style preferences, interaction history, implicit commitments, relationship dynamics — all retained so Arjun never re-introduces anyone.
Builds after hire
🎯
Leads & prospects
Qualification history, interaction log, next steps
🤝
Customers
Deal context, preferences, relationship health
🔗
Partners
Context, agreements, relationship dynamics
💭
Communication style
How each person prefers to be spoken with
System 6 · Proactive Intelligence Network (PIN)
Arjun watches specific signals — and briefs you before you ask
Event subscriptions, not cron polls. Arjun watches domain-specific signals that actually matter for their function. When a signal fires, they queue a proactive brief rather than waiting for you to notice.
Always watching
Arjun's 7 active watch patterns
WATCH
Average confidence score dropping below 80% across active intents (training data drift)
WATCH
Failed intent rate (unrecognized queries) climbing week-over-week
WATCH
Human handoff rate exceeding target (bot unable to resolve intended query types)
WATCH
Dead-end flow abandonment rate climbing (missing exit paths)
WATCH
Post-bot ticket volume increasing (resolution quality dropping)
WATCH
Intent regression after model update (existing functionality broken)
WATCH
Shadow test showing >15% divergence from human response quality
System 7 · Action Layer — Trust Ladder
Four autonomy modes — capabilities earn trust, not time
Arjun starts at Research Only. Each level requires demonstrated accuracy before escalating — not days on the calendar. You can also grant or revoke autonomy per-task type at any time.
Starts: Research Only
Research Only
Intent performance auditFailed query analysis and training data gap identification
Draft for Approval
Flow design and intent architectureTraining data sets and response variants
Act with Notification
Model updates within pre-approved intent scopeA/B test activation within configured parameters
Fully Autonomous
None by default — owner unlocks after track record demonstrated
System 8 · Meeting Intelligence Loop
Pre-brief → live notes → action items owned to completion
The gap no competitor fills. Most AI tools stop at the meeting. Arjun briefs you before, captures decisions during, extracts action items after, and follows each item to completion — no decisions lost, no follow-through broken.
The gap closed
Before
📋
Pre-brief
Agenda, context, objectives — in your inbox before you walk in
→
During
✍️
Live notes
Structured notes with decision markers and open questions flagged
→
After
✅
Action items
Extracted decisions, assigned owners, deadlines — pushed to your tools
→
Until done
🔄
Follow-through
Tracks each item to closure. Flags stalled items before they become forgotten commitments
System 9 · Outcome Attribution
Tracks what worked, what failed, and why — so mistakes don't repeat
Arjun owns their KPIs. Every outcome — good or bad — feeds back into their judgment. Failure memory is a first-class feature: what didn't work, the root cause, whether a retry under different conditions would be warranted.
Self-reporting
Arjun's 6 owned KPIs
KPI
Correct resolution rate per intent (primary — not just deflection)
KPI
Average confidence score across all active intents (target: >85%)
KPI
Human handoff rate (target vs. design)
KPI
Bot CSAT vs human CSAT on comparable query types
KPI
Failed intent rate (% of queries the bot cannot classify)
KPI
Time to first correct resolution (bot speed vs human baseline)
System 10 · Cross-Employee Cortex (CEC)
Persistent shared intelligence across every employee you hire
When Arjun discovers something that changes how the business should operate, that organizational intelligence is available to every other employee — without a meeting, without a memo, without anyone remembering to tell anyone.
Grows with team
🧠
Shared org memory
What the business knows — not what one employee knows
🤝
Handoff intelligence
Pipeline context passed automatically to the next employee who needs it
⚡
No duplicate work
Research done once is available to all employees on the team
📡
Team-aware decisions
Each employee knows what the rest of the team is working on