An internal AI copilot that helps TasteCraft's sales team respond faster, close more deals, and personalize every customer interaction.
TasteCraft is a premium ready-to-eat meal service based in Ontario, Canada, launching in Quebec in Q3 2026. We serve customers across four dietary tracks: Protein Forward, GLP-1 Friendly, High Fiber, and Vegetarian.
This AI copilot is NOT a customer-facing chatbot. It's a decision-support tool for our sales team (customer support reps, account managers, and sales development reps) who handle phone calls, emails, live chat, and follow-ups.
Primary users: TasteCraft sales team members
- Customer support reps (handle objections, refunds, delivery issues)
- Account managers (manage corporate clients and subscriptions)
- Sales development reps (convert leads from free trials to paid)
Secondary users: Team leads and trainers (use the copilot to onboard new reps)
- New reps take 2-3 weeks to memorize 40+ meals across 4 tracks
- Reps give inconsistent answers to the same customer objection ("Is GLP-1 safe?")
- Follow-ups are generic ("How was your meal?") instead of personalized
- Winter delivery complaints spike and reps don't have standardized protocols
- Quebec expansion (2026) requires training an entire team on new market nuances
- Reps search natural language: "Customer on Ozempic worried about portion size" → instant playbook
- Response times drop from 8 minutes to 2 minutes (estimated)
- First-call resolution improves (reps answer without escalating to managers)
- New reps are ramped in 3 days (vs 2-3 weeks) using the knowledge base
- Seasonal scenarios (cottage delivery, snowstorms, March break pauses) are pre-scripted
Your sales rep types a customer question into the AI copilot. The AI reads your product catalog, FAQs, objection scripts, and customer personas. It then outputs three things:
- SUGGESTED RESPONSE - What the rep can copy-paste to the customer
- INTERNAL NOTES - Private guidance for the rep (persona, next steps)
- FOLLOW-UP QUESTION - What to ask the customer next
The rep edits the response (adds personal touch) and sends it. Total time: ~90 seconds.
| Component | Tool | Why This Choice |
|---|---|---|
| Agent Framework | DIO Agent (deagent) | Lightweight, uses simple markdown files, no coding required |
| Runtime Environment | Google Antigravity | Free, browser-based, runs agents without local setup |
| AI Model (LLM) | Gemini 3.5 Flash | Free tier, fast responses, handles Canadian English well |
| Knowledge Base | 4 Markdown files | Simple RAG (Retrieval-Augmented Generation) without complex vector databases |
| Configuration | agents.md + cloud.md |
Defines behavior, response format, and escalation rules |
Step 1: Rep types a query
Example:
"Customer in Barrie, first-time buyer, worried about -20°C weather and food freezing on porch. No subscription. Help me respond."
Step 2: The agent reads the knowledge base
It searches these 4 files in priority order:
| Priority | File | What It Looks For |
|---|---|---|
| 1st | objections.md |
"Customer worried about winter delivery" |
| 2nd | personas.md |
"Barrie + first-time + no subscription" → Persona 5 |
| 3rd | faq.md |
"What happens if food freezes?" |
| 4th | product.md |
"What meal should I recommend?" |
Step 3: The agent applies rules from agents.md
These rules include:
- Always output 3 sections (SUGGESTED RESPONSE, INTERNAL NOTES, FOLLOW-UP)
- Never speak to customers directly (you are an internal tool)
- Prioritize Canadian context (Ontario winter protocols)
- Escalate to manager for medical or legal concerns
Step 4: The agent generates the response
The AI model (Gemini 3.5 Flash) combines:
- The customer query
- The relevant knowledge from your 4 files
- The formatting rules from
agents.md
Step 5: The rep receives a structured answer
## SUGGESTED RESPONSE
[Copy-paste this to the customer]
## INTERNAL NOTES (Rep Only)
- Persona: Persona 5
- Next action: Offer 3-meal trial, no subscription
## FOLLOW-UP QUESTION
[Open-ended question to ask customer]┌─────────────────────────────────────────────────────────────────┐
│ SALES REP (HUMAN) │
│ │
│ "Customer in Barrie worried about -20°C. Help me respond." │
│ │
└───────────────────────────────┬─────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ GOOGLE ANTIGRAVITY (Runtime) │
│ │
│ ┌────────────────────────────────────────────────────────────┐ │
│ │ DIO AGENT (deagent) │ │
│ │ │ │
│ │ 1. Reads `agents.md` for behavior rules │ │
│ │ 2. Reads `cloud.md` for runtime config │ │
│ │ 3. Loads 4 knowledge files into context │ │
│ │ 4. Calls Gemini 3.5 Flash API with prompt │ │
│ │ 5. Formats output as 3 sections │ │
│ │ │ │
│ └────────────────────────────────────────────────────────────┘ │
│ │
└───────────────────────────────┬─────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────┐
│ KNOWLEDGE BASE (Markdown Files) │
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ |
│ │ product.md │ │ faq.md │ │ objections.md│ │ personas.md │ |
│ │ │ │ │ │ │ │ │ |
│ │ 12 meals │ │ 35+ FAQs │ │ 20+ scripts │ │ 8 personas │ |
│ │ CAD pricing │ │ Delivery │ │ Price │ │ Tone guide │ |
│ │ Sourcing │ │ Winter │ │ Weather │ │ Follow-ups │ |
│ │ Seasonal │ │ Quebec │ │ GLP-1 │ │ Escalation │ |
│ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ |
│ │
└─────────────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ AI OUTPUT (Returned to Rep) │
│ │
│ ## SUGGESTED RESPONSE │
│ "I hear your concern - Barrie winters are no joke..." │
│ │
│ ## INTERNAL NOTES (Rep Only) │
│ - Persona: Persona 5 (Skeptical First-Timer) │
│ - Next action: Offer 3-meal trial, no subscription │
│ │
│ ## FOLLOW-UP QUESTION │
│ "What time works best for delivery?" │
│ │
└─────────────────────────────────────────────────────────────────┘
knowledge/
├── product.md # 12 meals, CAD pricing, Ontario sourcing, seasonal items
├── faq.md # 35+ FAQs (Ontario delivery, winter, Quebec expansion)
├── objections.md # 20+ objection scripts (price, weather, GLP-1, trust)
└── personas.md # 8 customer personas with AI tone guidelines
Total knowledge coverage: 8 personas × 4 tracks × 5 objection types × 3 seasons = ~480 scenarios
Sales rep types into AI copilot:
"Customer in Barrie, first-time buyer, worried about food freezing on their porch. It's February. They don't want to commit to subscription. What do I say?"
AI copilot output:
## Suggested Response (copy-paste to customer)
"I hear your concern - Barrie winters are no joke. Here's how we protect your food:
1. **Heated delivery bags** keep meals at 38-45°C for 2 hours in -20°C
2. **Text alerts** when driver is 5 minutes away - you can meet them at the door
3. **If food arrives cold** - full refund, no questions asked
We delivered all last winter during polar vortex days. Want to try 3 meals with 25% off? No subscription required."
## Internal Notes for Rep
- This customer is likely **Persona 5 (The Skeptical First-Timer - Winter)**
- Key objection: Trust + weather, not price
- Don't push subscription on first call
- Offer "weather guarantee" explicitly (full refund if cold)
## Follow-up Question to Ask Customer
"What time of day works best for delivery? We can aim for your lunch break (12-1pm) so you're home."Result: Rep copies the response, adds personal touch, sends in 90 seconds vs 8 minutes of research.
| Test | Topic | Status |
|---|---|---|
| 1 | Winter Delivery (Barrie) | ✅ PASS |
| 2 | GLP-1 Nausea | ✅ PASS |
| 3 | Price Objection | ✅ PASS |
| 4 | Quebec Launch (French) | ✅ PASS |
| 5 | Severe Allergy | ✅ PASS (safe refusal) |
| 6 | GLP-1 Fatigue (Mounjaro) | ✅ PASS (exceptional) |
| 7 | 5-Day Meal Planning | ✅ PASS (exceptional) |
Pass rate: 7/7 (100%)
Full test results here!
Click each test to see the AI copilot in action.
Test 1: Winter Delivery (Barrie, -20°C)
Customer worried about frozen food in Barrie winter. Agent responds with heated bag guarantee, weather guarantee, and no subscription pressure.
Test 2: GLP-1 Nausea (Ozempic)
Customer on Ozempic with nausea. Agent validates experience, mentions Dr. Sarah Chen at Toronto General Hospital, recommends Gentle Roots Bowl with ginger-turmeric broth.
Test 3: Price Objection (Tim Hortons Comparison)
Customer says "$14 for lunch? I can get Tim Hortons for $6." Agent anchors to premium competitors ($15-18), highlights Ontario sourcing, offers 25% off.
Test 4: Quebec Launch (French)
Customer in Montreal asks in French about Plateau delivery. Agent responds in natural Quebec French, confirms September 15 launch, offers waitlist with 30% discount.
Test 5: Severe Peanut Allergy
Customer with severe peanut allergy. Agent safely refuses to invent information, escalates to manager, does NOT guarantee safety (correct behavior).
Test 6: GLP-1 Fatigue (Mounjaro)
Customer on Mounjaro with nausea and fatigue. Agent generalizes correctly, addressing fatigue even though not explicitly in knowledge base.
Test 7: 5-Day Meal Planning
Customer asks for 5 days of lunch + dinner. Agent recommends Standard Plan (10 meals), structures lunches (High Fiber for energy) vs dinners (Protein Forward for recovery), provides 6 specific meals with nutritional reasoning.
- Google account (free)
- Antigravity CLI (free)
# Install Antigravity CLI
irm https://antigravity.google/cli/install.ps1 | iex
# Run the agent
agy --help| Metric | Before Copilot | After Copilot (Target) |
|---|---|---|
| New rep ramp time | 2-3 weeks | 3 days |
| Response to common objections | 5-8 minutes | 90 seconds |
| First-call resolution rate | 68% | 85% |
| Customer satisfaction (CSAT) | 4.2 / 5 | 4.6 / 5 |
| Winter delivery refund rate | 12% | 6% (better upfront communication) |
| Quebec launch readiness | 2 months training | 2 weeks (knowledge base translated) |
-
Quebec French version: Translate all 4 knowledge files to French, add Quebec-specific personas (Montreal plateau, Quebec City families)
-
Integration with CRM: Auto-populate customer persona based on order history (e.g., "This customer ordered GLP-1 3x - use gentle tone")
-
Response analytics: Track which suggested responses get used most, optimize the knowledge base
-
Real-time objection handling: Integrate with call software to transcribe customer objections and surface responses live (like Gong.io but for sales)
-
Predictive follow-ups: AI suggests "This customer will likely cancel in 2 weeks (no order in 10 days) - send win-back offer now"
-
Sales training simulator: New reps practice on AI-generated customer personas (using our 8 personas as test cases)
-
Fully autonomous agent: For low-complexity issues (delivery rescheduling, payment updates), AI handles directly. Human rep only for high-value or complex cases.
-
Cross-sell recommendations: "Customer ordering Protein Forward at 3pm daily - recommend High Fiber for breakfast add-on"
DIO-Project-AI_Sales_Copilot
├── examples
│ ├── chat_consumer_price_objection.md
│ ├── chat_consumer_question_on_freshness.md
│ ├── follow-up_after_sale.md
│ └── test-results.md # 7/7 tests passed
├── knowledge
│ ├── faq.md # 35+ FAQs (winter, Quebec, delivery)
│ ├── objections.md # 20+ objection scripts
│ ├── personas.md # 8 customer personas
│ └── product.md # 12 meals, CAD pricing, Ontario sourcing
├── prompts
│ ├── consumer_analysis.md
│ └── copilot_answers.md
├── assets/
│ ├── cli-demo-q1.png
│ ├── cli-demo-q2.png
│ ├── cli-demo-q3.png
│ ├── cli-demo-q4.png
│ ├── cli-demo-q5.png
│ ├── cli-demo-q6.png
│ └── cli-demo-q7.png
├── agents.md # Agent behavior rules
├── cloud.md # Runtime configuration
├── Company_Identity.md # Brand guidelines
├── LICENSE
└── README.md # Complete project documentation
-
Localization matters: An Ontario customer in February has different needs than a Quebec customer in September
-
Dietary tracks reduce friction: Customers don't want to calculate macros - they want "GLP-1 friendly" as a simple label
-
Sales copilots > chatbots for complex products: Ready-to-eat meals have too many exceptions (weather, allergies, delivery zones) for fully automated customer support
-
RAG doesn't need vector databases for small domains: 4 markdown files work perfectly for <500 scenarios
-
Persona-driven prompts improve response quality dramatically (comparing generic vs persona-specific outputs)
-
Tiered response structure (suggested response + internal notes + follow-up) balances automation with human control
- Inspiration: DIO's "Copiloto de Vendas com IA" challenge
- Reference data: Sweetgreen, Chipotle, and Canadian meal services (Fresh Prep, Goodfood, Factor, CookUnit)
- Ontario farm partners: (fictional but based on real farms: Cumbrae's, Hewitt's, Forge & Flour)
This is a portfolio project for Raquel Marques. For questions or feedback:
- GitHub Issues
- LinkedIn: Raquel Marques
- License: MIT (feel free to fork and adapt for your own portfolio)
- Last updated: June 2026
This agent is compatible with DIO Agent (deagent) running on Google Antigravity with Gemini 3.5 Flash (free tier).
agents.md- Agent behavior, response format, escalation rulescloud.md- Runtime configuration for Antigravity + Gemini
Prerequisites:
- Google account (free)
- GitHub account (free)
Step 1: Import to Google Antigravity
- Go to https://antigravity.google
- Sign in with Google account
- Click "Import from GitHub"
- Select
DIO-Project-AI_Sales_Copilotrepository
Step 2: Get Gemini API key (free)
- Go to https://aistudio.google.com
- Click "Get API key"
- Create key (no billing required)
- Copy the key
Step 3: Add API key to Antigravity
- In Antigravity, go to Settings -> Secrets
- Add
GEMINI_API_KEYwith your key
Step 4: Run the agent
- In Antigravity terminal, run:
deagent run --config agents.md --runtime antigravityStep 5: Test with a query Type: Customer in Barrie worried about -20°C weather. Help me respond.
Expected Output The agent will return three sections:
- SUGGESTED RESPONSE - Copy-paste to customer
- INTERNAL NOTES - For rep only (persona, next steps)
- FOLLOW-UP QUESTION - Open-ended question to ask customer
