My production runs on an Azure VM with Docker Compose.
My production is deployed by GitHub Actions on every push to main (deploy.sh).
My production domain is https://plantaeai.tech.
My web container (plantae) runs migrations, setup_checkpointer, collectstatic, then Gunicorn.
My worker container (plantae-worker) runs python manage.py qcluster.
My worker handles Slack notifications, AI replies after staff decisions, and SLA reminders.
My static and media files are collected into static/ and served by Apache.
My chat widget's CSS/JS URLs carry CHAT_WIDGET_VERSION, so browsers fetch new versions after a deploy.
My Azure Database for PostgreSQL also stores agent memory and the background task queue.
My initial configuration step is Copy .env-sample to .env.
My Slack configuration uses SLACK_BOT_TOKEN, SLACK_SIGNING_SECRET, SLACK_ESCALATION_CHANNEL.
My Slack app credentials and escalation channel are for tickets.
My HITL_STAFF_EMAILS setting is a comma-separated list for email fallback and overdue reminders.
My reminder and expiry settings are HITL_SLA_MINUTES, HITL_EXPIRE_HOURS.
My coupon configuration is Coupon limits (default 2 units, 7 days).
My SITE_URL setting is the public URL used in links from Slack/email.
My SITE_URL is the public URL used in links from Slack/email.
My widget asset update rule is Bump when the chat widget's CSS/JS change.
My database setting is Database SSL mode (default require).
My Slack app bot scopes are chat:write, users:read, users:read.email, channels:history (and groups:history for a private channel).
I install the Slack app and invite it to the escalation channel.
My Slack setup uses the Interactivity request URL https://<your-domain>/agent/slack/interactions/ and the Event Subscriptions request URL https://<your-domain>/agent/slack/events/.
My Slack app subscribes to message.channels or message.groups.
My contact email is ayush.ttps@gmail.com.
I match staff to site accounts by email and give their accounts is_staff so decisions are recorded under the staff member’s name. If Slack is being configured, the source uses <your-domain> as a placeholder in the request URLs rather than specifying the actual domain. The source does not provide values for the Slack credentials, escalation channel, staff email list, SLA or expiry timings, coupon overrides, or widget version.
My project is licensed under the MIT License.
My contribution guidance is: Pull requests are welcome! Please open an issue first to discuss changes.
I performed load testing with Locust using the scenario in locust_loadtest/locustfile.py; the recorded scenario value is Locust using locust_loadtest/locustfile.py.
My production suite mixes three weighted user types: anonymous browsers (weight 4), authenticated customers who log in and exercise cart/checkout paths (weight 3), and a cart-profiling user that hammers add_cart (weight 1).
Each page load fetches its referenced static and media assets, mirroring real browser traffic.
I left order-placement POSTs disabled in both runs so no test orders were written to the production database.
My baseline run used 50 concurrent users, spawn rate 5/s, 3-minute run.
My baseline run recorded 8,351 total requests; 5 total failures; 0.06% failure rate; 46.7 requests/second.
My baseline aggregate response times were 44 ms median; 3,300 ms 95th percentile; 6,200 ms 99th percentile; 7,597 ms maximum.
The baseline included ~7,400 static/media requests; Apache served them in 42-43 ms.
Baseline static and media assets had a 95th percentile between 55 and 360 ms; zero failures.
Baseline browse pages had 2,800-3,000 ms median; dashboard and my-orders had 2,600-2,700 ms; login POST had 2,000 ms.
Baseline cart-add latency was 6,000-6,200 ms median; roughly 2x any read path.
The baseline failures were four RemoteDisconnected errors; one 403 from a stale CSRF token on a retried cart-add POST.
My stress run used 200 concurrent users, spawn rate 10/s, 3-minute run.
My stress run recorded 11,209 total requests; 591 total failures; 5.27% failure rate; 62.6 requests/second.
Stress-run aggregate response times were 52 ms median; 12,000 ms 90th percentile; 19,000 ms 99th percentile; 30,880 ms maximum.
During the stress run, dynamic requests had ~13,000 ms median across every browse and account view; 25,000-31,000 ms median on add_cart POSTs.
Stress-run authentication and CSRF failures included 23 failed logins; 10 unreachable dashboards; CSRF-dependent cart POSTs 403'd because the page that carries the token never loaded.
My capacity sweep used Anonymous browse journey only (no login), 90 s per level, spawn rate 20/s.
The anonymous sweep recorded 100: 0.43% failure, 99.57% success; 150: 0.63%, 99.37%; 200: 17.24%, 82.76%; 250: 23.03%, 76.97%; 500: 27.28%, 72.72%.
The healthy-service ceiling was ~150 concurrent unauthenticated users; >99% success.
My test results indicate capacity of ~150 concurrent unauthenticated browsers or ~50 concurrent authenticated customers before failures climb.
My measured dynamic Django throughput is capped at ~15-20 requests/second.
The observed dynamic-request limit is consistent with a small Gunicorn worker pool where each database-writing request holds a worker for seconds.
My load-test analysis concludes: The bottleneck is not the network: static assets from the same host returned in ~43 ms across every run.
My anonymous browse journey covered GET /, /store/, /store/category/plants/, /store/search/, and product detail pages.
My authenticated journey covered login, dashboard, my orders, browse, POST /cart/add_cart/{id}/, cart view, and checkout page.
My cart-profiling journey repeatedly posted to POST /cart/add_cart/{id}/ and read the cart and checkout pages.
My load test analysis found that quadrupling users barely moved throughput, from 46.7 to 62.6 req/s, while latency and errors collapsed.
During the stress run, responses capped near 31 s, indicating a proxy/Gunicorn timeout ceiling rather than eventual completion.
In the stress run, most failures were RemoteDisconnected as the server dropped connections.
Between 150 and 200 users, the site fell off a cliff as Gunicorn's queue overflowed and connections were dropped.
Past 250 users, throughput drops while errors climb; beyond the tested limits, added concurrency produces queueing and timeouts, not more work done.
My raw Locust result files are in locust_loadtest/ (plantae_loadtest_* for the baseline, plantae_stress_* for the stress run).
My full HTML reports are at plantae_loadtest_report.html and plantae_stress_report.html.
For proposal-based escalations, I handle these cases by collecting details, checking them against the database, and proposing an action: Price match / better deal, order cancellation, delivery issues, payment issues.
I hand these cases over for direct staff conversation: Damaged or wrong item, bulk orders, complaints, asks for a human, or clearly upset.
Customers can use the 👤 Talk to a human button or suggestion in the chat at any time.
My Slack ticket card contains a summary, customer's details, facts from the database, and the AI's proposal.
When staff approve a proposal, the action is carried out and the AI tells the customer.
For an edit-and-approve example, staff change the proposal from 15% → 10% and add a note; the AI tells the customer the edited outcome.
My rejection workflow is: Staff give a reason; the AI explains it to the customer; nothing is changed.
In the Slack ticket, Take over chat switches to a live conversation.
When staff take over, the AI goes quiet.
During takeover, Customer messages appear in the ticket's Slack thread, and staff replies in that thread appear in the customer's chat labelled with the staff member's name.
My handback behavior is: Hand back to AI closes the ticket; the AI then remembers what staff said.
My ticket timing settings are: HITL_SLA_MINUTES reminder threshold; HITL_EXPIRE_HOURS email-follow-up notice. A reminder is posted after the threshold, and the customer is told the team will follow up by email after the expiry setting.
If Slack is unavailable, staff are emailed and can decide everything from Django admin Escalation Tickets with the same Approve / Edit / Reject / Take over / Reply actions.
My ticket audit record is: Every action is recorded on the ticket (who, when, what), visible in the admin.
An approved price match can issue a customer-specific coupon code such as ROSE10; another customer entering it gets “not valid for your account”.
My price-match coupon restriction is: Only for that product.
The coupon limit is up to 2 units per order (PRICE_MATCH_MAX_UNITS).
My coupon validity and use limit is: Valid for 7 days (PRICE_MATCH_VALID_DAYS) and single use.
The coupon is marked used only after Razorpay confirms the payment.
Customers enter the coupon in the coupon box on the cart or checkout page.
My coupon tax treatment is: discount is shown before tax; 18% GST is charged on the discounted amount and the discount is saved on the order.
My admin coupon record fields are owner, product, %, expiry, when used, order, and the ticket that created it.
For price matches, cancellations, delivery issues, and payment issues, I collect the details (product, price, order number), check them against the database, and propose an action for staff to approve.
My escalation tickets wait in Slack until staff act.
My Slack edit-and-approve screenshot is docs/screenshots/slack_edit_approve.png.
Another caption for my Slack screenshot is “Edit & approve: the final price updates as the discount changes”.
I built Plantae 🌱, an AI-powered e-commerce platform for plants and gardening products, in a repository on my GitHub profile, my github profile. Its production site is https://plantaeai.tech.
My Plantae repository metadata lists HTML, 1 fork.
My Plantae repository topics are azure, css, django, gpt5, html, js, langchain, langgraph, openai, postgresql, python.
My Plantae repository was started 2025-07 and last updated 2026-10.
My project is a Django application deployed on an Azure VM at plantaeai.tech.
My Plantae production URL is https://plantaeai.tech.
My Plantae assistant supports plant care queries, product recommendations, order status, and cart management, including quantity requests such as “add 2 marigolds”.
My Plantae assistant supports text and image-based queries; can identify plants from images.
My Plantae assistant understands casual and mixed-language messages (English/Hindi/Hinglish, typos).
My Plantae assistant politely declines off-topic requests and attempts to override its instructions.
My Plantae assistant Remembers the conversation across sessions (stored in PostgreSQL).
My Plantae assistant's rate limit is max 10 AI messages per user; messages to staff during a takeover don't count.
My Plantae assistant provides Text-to-Speech (TTS) and Speech-to-Text (STT) using ElevenLabs.
My Plantae Supervisor Agent Routes each message to one sub-agent using LLM-based classification (with recent conversation as context).
My Plantae Cart Agent supports Add/view/remove items in cart. I use separate Cart, Order, Research, Recommendation, General, and Escalation Intake modules.
My Plantae cart workflow uses Variation selection only for cart; when a variation is needed, the flow goes through Variation Selection and returns to the Cart Agent.
My Plantae image workflow follows the rule: Plant identification augments user input if image is uploaded.
My Plantae Escalation Intake collects details staff need and hands the request to a human; tickets go to Slack: Plantae team.
In my Plantae workflow, Agent responses pass through a Response Node and are returned to the user.
The Order Agent fetches order details and order history, and redirects customers to checkout and their orders. The Research Agent handles plant care, diseases, watering, and sunlight; the Recommendation Agent suggests products based on user needs and the catalog. The General Agent handles greetings, thanks, and questions about the conversation. Escalation Intake gathers information staff need and hands the request to a person. My detailed agent and escalation architecture is documented in agent/README.md.
The store combines account and catalog features with cart and checkout, orders, reviews, and administration.
My Plantae authentication features are Register, login, logout, email verification, password reset, profile management.
My Plantae catalog lets customers Browse, search, and filter products by category, price, and keyword.
My Plantae product pages provide Detailed product pages with images, plant care info, reviews, and variations.
My Plantae cart and checkout support Add/remove products, manage variations, view cart, apply user-specific coupon codes, checkout with tax calculation.
My Plantae order management supports Place orders, view order history, order details, payment via Razorpay, order confirmation emails.
My Plantae review feature lets customers submit and update product reviews.
My Plantae admin panel can Manage users, products, categories, orders, and chat limits.
My Plantae interface uses responsive design with Bootstrap and custom CSS.
My Plantae assistant supports Plant care, product help, and order support via text, image, and voice, with starter suggestions.
My Plantae conversational agent supports plant care, product recommendations, and order support.
My Plantae human-in-the-loop workflow is: Requests the AI can't handle (price matches, cancellations, refunds, damaged items, bulk orders, complaints) are escalated to the team on Slack, where staff approve, edit, reject, or take over the chat live.
My Plantae price-match policy is: An approved price match creates a coupon code that only works for that customer and product.
The README includes screenshot sections for the home page, admin, store, product detail, cart, payment success, dashboard, coupon at checkout, chat widget, and chat suggestions. It references Slack screenshots in the Human-in-the-Loop section; blank screenshot-table cells do not establish whether other images exist.
My Plantae logo image is specified with width 180.
My Plantae dashboard screenshot is docs/screenshots/dashboard.png, specified with width 420.
My Plantae coupon-at-checkout screenshot is docs/screenshots/cart_coupon.png, specified with width 420.
The README does not give implementation details for the AI model, chat-limit policy, tax calculation, Razorpay integration, Slack integration, or price-match approval process. It also does not explain the 2026-10 last-updated date or the meaning of the gpt5 repository topic.