09 · Final Round Whimsical: Transcription

info · Transcribed from your two large Whimsical boards: "delivery agent simulations brainstorming and changes to make for final version" (Board 1) and the big questions / madam otp / normal flow / return / agent-human flow / stages board (Board 2). Text is kept as written. […] marks a place where the board image is cut off at a seam or the text was too small to read with confidence; those spots are listed in 00 README Index so you can fill them from the live Whimsical.

BOARD 1 · "delivery agent simulations brainstorming and changes to make for final version"

1.1 Round 3 brief (pasted on the board, verbatim)

n Round 2 you designed your agent on paper. This round has two parts. First you run a simulation of your agent. Then you work out which company turns it into a business, and how. You'll build the simulation, using a method called Wizard of Oz. The parts that can run, run. The parts that don't exist yet are played by someone behind a curtain. Your agent is the part that runs. It's an AI model like Claude or Gemini, following a system prompt you write, and it makes every decision on its own. Wherever the model can connect to something directly, like Gmail, WhatsApp or Google Sheets, connect it. Everything else is played by a team member behind the curtain:

What your agent connects to - Gnani: run the audio or text through Gnani's API yourself, and paste back exactly what Gnani returns.. - Pine Labs and Delhivery: reply with the response their documentation says they'd send. - Anything else from the outside world, like a voice note arriving or the clock hitting 5 PM: a team member feeds it in. It must come from a real source. If no real source produces it today, your agent can't have it. The person behind the curtain never helps the agent think. They only play the world around it. If your agent needs someone to tell it what to do next, it isn't an agent yet.

Recording your simulation - Record your screen while your agent runs from the first thing that happens to the moment your agent's outcome is achieved. - The recording must show every input arriving, every call your agent makes to Gnani, Pine Labs or Delhivery and the response that comes back, and every message your agent sends to a person. - Keep it to 5 minutes at most. Fast-forward the parts where the model is thinking. - Share the link, along with your system prompt and the name of the model you used.

Part 1 (mind map, verbatim)

• part 1 1. Tell the story of one person using your agent. Name them, name the date, and tell it in the order it happens, the way you'd explain it to a friend. Say what they see, what they say, and what your agent does in reply. 100 words at most.
• Then list every decision your agent makes in the recording, in order. A decision is any point where your agent picks one thing over another. For each one, give: 1. What does your agent know on day one, before anyone has used it? Who gives it that information, and how? 1. For every call your agent makes to Pine Labs, Delhivery or Gnani, name the exact endpoint from their documentation, and share the request and response. You can imagine up to 3 capabilities that don't exist yet. List them separately. For each one, name the partner, write the endpoint, and tell us what data that partner already holds that makes it possible. 1. Show us a mockup of every screen or conversation where your agent meets a person, whether that's a WhatsApp chat, a call script, an app screen or a notification. 1. Score each rail out of 10 on how agent-ready it is, and give one reason for each score.

Part 2 (mind map, verbatim)

• part 2 — "If your agent works, which company turns it into a business, and how?"
• Where does your agent live, and how does a customer start using it?
  • Is it its own app, a feature inside another company's app (name the app), a WhatsApp number, or something else? → your note: whatsapp number, calls, ekart wa business chat
• Which company should build it, and why them?
  • What do they already have that nobody else does?
  • Why haven't they built it yet?
• Who will they sell it to, and how will they make money from it?
  • What do those customers do today instead, and what does that cost them?
  • Will the customer pay, or will the company earn from something else, like more orders? → your notes: "[…] ask for a small payment like how we d[…] ut - for a seamless experience (exp1)" / "How to convince on getting more revenue since delayed orders % will go down and increasing operational efficiency and unit economics"
• How will the agent get its first 1,000 users? This is the question we care about most, so be as inventive as you like.
  • Where do these people already spend time, and what gets someone to try it the first time? → your note: "order not delivered and DA marked" customer available "can be a good trigger . what othe[…]"
  • What keeps them using it after three weeks?
• Make the landing page for the agent the way the company you named in question 2 would, with a call to action to sign up, and share the link.
• What could kill this agent in its first year?
  • regulations especially for agentic COD
  • false negatives → if the voice agentts confidence scoring is even a little off, it starts marking people as unavailable who were actually home, which pushes rto up instead of down,
  • if the entire agent fails at scale
  • internal roadmap cans shift. people can deprioritise the problem if the delta between this agentic solution ain't big enough as to the existing solution of just using OTP (since OTP behavior has already been set) → whoever owns this thing across teams can easily get deprioritized the moment theres a reorg or someones roadmap shifts, and a good pilot with no clear long term owner
  • if the TOKEN cost for gnani and other models increases and the unit economics doesn't make sense to add this agent (kabootar). since deliverys are done at very high scale and even a marginal increase of token price and other → counter: although a counter argument could be by replacing the API endpoints and using open source models instead of proprietary models
• assumption (yellow): lets assume the onboarding process of this agent(kabootar) through marketing and traning of delivery agent is done properly
• (pink, repeated at bottom of board) Where do these people already spend time, and what gets someone to try it the first time? / What keeps them using it after three weeks?

1.2 Simulation screens and the changes you wanted for the final version

These are the screens from the simulation tool with your handwritten change notes next to each. Screens are described only so your notes make sense; the notes are verbatim.

1.3 WhatsApp reference screenshots

• Label on board (red): "reference image and screen shots of WA chat for flipkart/ekart"
• Label on board (yellow): "flipkart has flipkart sathi which their own whatsapp chat to send messages"
• Shadowfax Delivery WhatsApp thread (reference): "Hi Paramesh, Your Mobfest Anti-T.. from Flipkart is out for delivery today! Tracking ID … To receive it, share the OTP - 6958 with the rider. Contact rider … Track or raise concern … To help us deliver smoothly, please select suitably" with buttons Yes, I am Available / No, I am not Available / Pay via UPI, a "Send Location" prompt, and a delivered message with Excellent / Average / Poor rating buttons.
• Flipkart Saathi thread (reference, 17 September 2026): "Confirm your availability for Pickup: Paragon Lightweight Stylish… For a smooth pickup experience, confirm your availability for the order pickup today between 7 am - 11 pm. Contact Delivery Partner … To pause important order updates via WhatsApp, send STOP." with buttons Available for the pickup / Reschedule the pickup.

1.4 Flipkart mind map ("why flipkart") and sources

• flipkart → logistics companies acquired so far · supermoney · ekart logistics
• why flipkart: this agent needs 3 rails and flipkart has all these 3: supermoney, voice agent (ring ai), and ekart · deelivery infrastructure with live location data, a payments rail that can hold and capture funds programmatically, and a voice/conversational layer that can talk to a buyer in their language. · flipkart is the only indian company sitting on all three. ekart for logistics, super.money for the UPI reserve pay and capture mechanics, and SLAP/their in house conversational team for the voice muscle along with partnership with ring ai
• my best guess on why they haven't build yet: i think fundamentally if you think what could be the internal north star or metric set up for ekart, supermoney and voice agents · ekart: cost per delivery or RTO %. this has notthing do with OTP friction · supermoney : TPV or maybe app rank and most likely not delivery conversions · plus maybe flipkart would want to try for IPO and increase their market capturue in QC · so i would say flipkart just haven' given this a lot of thought since they are super busy with their current priorities · seconly no one is the org would bonus in building this. this feature/ stack could be deprioritized even if someone did bring it up · or they are actually building this but deploying at low scale. maybe to the top 20-30 pin codes which gives them a huge chunk of their GMV
• sources (as listed on the board): Super.money launches AI shopping agents in India | The Paypers · How RazorpayX Enabled Super.money to Deliver Instant Cashbacks - Razorpay Blog · Ekart Tries AI for Return Delivery Calls in India | Sushant Shekhar posted on the topic | LinkedIn · AI Voice & Chat Agent Platform for Businesses | Ringg AI

1.5 How will you monetize this agent (mind map)

• What do those customers do today instead, and what does that cost them? / Will the customer pay, or will the company earn from something else, like more orders?
• how will you monetize this agent
• add on feature to make the "experience better" → increase unit econonmic and operational efficencey
• premium feature for 1. flipkart black 2. flipkart VIP 3. flipkart plus gold → positioning
• just like how you pay "extra 15-25 ruppes" for COD, can we pay extra for such orders?
• umm can we also sell this to other D2C brands? who bear the cost of RTO
• can we license these 3 rails to other players? → will this open up new revenue streams? but does flip[kart] has a bandwhith to do this right now?
• what is the pain point (cost) → repetitive: useless task for people who have high $/hr · frustating: order getting rescheduled the next day is frustr[ating]
• saving money internally, RTO on cod orders are 15-20%, and this cost money to the seller.
• 2) every time the delivery agent is spending time for getting the OTP, he is not deliverying at 30 other stops. → second order effect: this will drive more orders p[er] place on flipkart therefore increasng their GMV
• so what total money saved through lower % OF RTO and more delivery/ rider
• ekart can do for other business, d2c brands. it can priced as a small premium on top of regular cod. → target brands who are loosing money to high RT[O]
• license the entire thing as an individual stack and sell it to other logistic companies. → aws model

1.6 Where does your agent live, and how does a customer start using it?

• buyer → whatsapp number: through which the agent 1) sends messages 2) calls (voice agent) → the buyer/ alternative reciever is alre[ady …] get updates about the delivery
• delivery agent → a layer in the existing EKART interface that the delivery agent see → he already open[s …]
• for COD → it can be a toggle in existing flipkart app to enable "agentic COD using super money

1.7 How will the agent get its first 1,000 users?

• first pick the 20% pincodes that drive 60-70% of the GMV for flipkart
• pick 10 high extreme dense pin codes (indiranagar, kormangla, bandra)
• for agentic COD → Roll out on invite basis for cohorts with high frequency of COD order and who have high RTO rate. · they are core users and this feature can push them to power users cohort
• whats getting someone to try it the first time, nothing really, theres no opt in. its just the new default. the otp step is gone, and since youre removing a chore instead of asking someone to make a decision, youll probably get close to 90% percent adoption on day one without spending a rupee on marketing.
• whats keeping them around after three week → buyers: its just a few deliveries quietly going right, especially that moment where mom picks it up because the whatsapp message caught it early instead of a missed delivery and an annoying reschedule call. · delvery agent: money and time, less standing around at doors, fewer failed drops he has to anyways redo later.

BOARD 2 · Questions, "madam otp", normal flow, return, agent-human flow, stages

2.1 Round 2 questions with your answers on the map (the "questions" tree)

Pasted brief (verbatim): "Design an agent that is specified clearly enough for someone else to build. The submission should clearly define: 1. What the agent is accountable for. 2. How autonomous the agent is. 3. How the agent behaves across different states and scenarios. 4. What existing capabilities the agent can use. 5. What capabilities are missing and need to be built. 6. How humans interact with the agent. 7. What the agent is called. 8. Which Indian company could potentially build an agent like this."

8. Which Indian company has the best chance of creating an agent just like yours? → flipkart and amazon · they own everything from sku to in house delivery service

1. Agent Outcome — What is the outcome your agent is accountable for? - voice agent → outcome → Understanding the context of the end user: whether he is available or not available · whether his family members are picking up · whether he wants to reschedule itand then giving that information to the delivery agent → to give a live status check if the rider s[…] that address or move to the next addre[ss] - logistic agent → delivery → outcome → removing the friction caused by OTP · returns → outcome - payment → outcome → Automatic payment for COD when the buyer is not available

2. Level of Autonomy — L1 Prepares things for the person. They still do the work. · L2 Proposes a specific action every time. The person says yes or no to each one. · L3 Acts on its own inside limits somebody set. The person deals with whatever falls outside those limits. · L4 Plans several steps, does them, checks its own work, comes back when stuck, and is judged on whether the thing got done.

3. Agent States and Flows — happy flow (cell connectivity? loud noise?) · unhappy flow (delivery agents camera is broken · end users phone is cracked)

4. Existing Rails and Missing Capabilities → can we have a call id display (delhivery/ delivery agent)

5. Fourth Rail → "what is a solution that makes it an innovation when compared to the existing technology ??" - 1) assuming otp is the problem and cause of friction - attach a camera to the delivery guy to remove the otp in intself , so assuming we can run small language models on edge devices , the computer vision software will be able to run and extract features real time and analyze which parcel is sent to whome without the otp - as the otp here solves the authentication problem which is more authentic than giving it to the same guy that ordered the delivery ? - 2)anytime anywhere and dynamic delivery application that analyzes your location real time and recommends if the delivery must be done to a specific location you are in or not - 3)telecommunication spam detection ? - 4)consumer ai which understands the users preferences and location and choices , for better machine to machine translation - (also on board) "this is why the 4th rail is the consumer ai / personal assistant"

6. Human Interface — (the seven prompts: how the user initiates interaction · how the agent communicates with the user · what information the user sees · when the agent asks for approval · when the agent acts independently · how the user handles exceptions or escalations · what the overall interaction looks like from beginning to end)

7. Agent Name — "Provide the name of the agent."

2.2 "questions to clarify" (your list, with the answers you reached)

details embedded in the qr → "QR - Qr will work even when the screen is shattered or during gl[are / bad] weather condition - So we will use pure and black QR · increase the padding between the UI and QR for better distinguis[hing] · Lets say if the users phone has that green line on his phone, still [works] since it will have maximum contrast · even if the screen is crack then make modules of the QR bigger · we will use Level H QR. to maximize error correction"

2.3 "madam otp" map (extended version)

Same branches as Map D in 03 Whimsicals Mind Maps, with these additions on the "exact ETA" and "multiple orders" nodes:

• exact ETA → tracking live gps is hard and expensive → "what if we can trigger a message of ETA by just calculating the last order before it reaches the next user and in that 10-15 mins window we schedule a message/ call" 1. available - if available then schedule the call to the priority list. 2. leave at the door - NO qr for this. - activate GPS, a proof of video - taking accountabilty usually decreases fraud and theft 3. not avaiable
• for multiple orders juggling between multiple otp is also irritating → single QR for multiple delivery - barcode of each order will be different - similar to how metro qr scan works
• otp → OTP is shared during the time of delivery - DA schedules it instantly · OTP is shared 4-5hrs before the agent reaches

2.4 "why returns" and "how to predict the right ETA?"

• why returns → mismatch product expectations and the reality → size is wrong · color is wrong · the item is only wrong · its broken
• how to predict the right ETA?
• probelms: live tracking also gives a false perception to the end user that the rider is nearby etc which will disappoint him if he doesn't come on times · live tracking is hard since a DA does 100-150 deliveries every day and tracking that live in a huge infrastructure cost
• assumption 1: delivery routes are pre determined and given the DA at the start of the day
• sol: what if we can send a voice call right after the previous delivery recorded since the DA will either successfully delivery or not delivery and that could be a trigger · for eg: if the agent finishes a successful delivery of an ORDER in A block of prestige apt then the next user will get a voice call that "hey our delivery agent is near, will come in 10-15 mins or "the DA is heading to your order
  • limitation → still we aren't improving the experience significantly · we are just giving a max 15-20 min prior update. → is that enough?????
• what can we do to give an update 1hr before or 3hrs before → "[…] calculate the t-10 order and then since there are 10 more orders and each order takes 10mins on avg then we just do 10*10 which is 100mins and then we give a time estimate. simply based on logic system."
• how is data segregated in these big companies? like delhivery

2.5 "4th rail" map

• delivery issue is a live data and context mismatch problem → the delivery agent doesn't know what is the current status of the end buyer · 1.if he is busy? in a meeting · 2)where exactly his location is
• "lets say delivery agent hypothetically knows the current information about the end buyer who placed. he knows everything in terms of who is available at his house, if he is a meeting, if he is at this office at scheduled delivery time, if he is close to his neighbor (who can then be trusted to delivery the package)" → india has a cultural habit of having some sort of assistance in daily life who makes low value and menial task decisions on behalf of us
• all of this information is the Alpha
• where does this information lies? in his phone right
• his whatsapp messages, his gps current location, if his phone is on DND then he is in a meeting and his location is office and his google calendar confirms that
• what would this rail do: this rail will communicate with logistic and voice agents through API and send live information about the current status of the end buyer who's parcel must be delivered
• it will send information such as: buyers live location/ busy not busy · other trusted people who can take the parcel on behalf of him · a message from the other person saying that he received the order · therefore confirming the actually delivery did happened and no fraud case
• all of this can be solved using personal assistant AI - consumer AI → "this lives on the users phone. Eg: with i[os] 27 update siri has the contextual understanding of the current state of the user"
• this is why the 4th rail is the consumer ai / personal assistant
• which Indian company will build this → cred really understands the user · zerodha does a lot of open source projects that can ma[ke] one of such personal bots · swiggy build a concierge agent · "currently idts any Indian company is building a personalized AI assistance. muse, instinct, poke are all western companies building such rail"

2.6 Normal flow of delivery (day by day)

on 1st , user placed the order (delivery is scheduled on 4th day ) (protien powder from flipkart ) 2nd day shipped to the whearhouse locally 3rd is dispatch 4rth morning is out for delivery and the driver is assigned by end of the day driver will deliver the order , driver will stack the order based on the optimised route by the application driver will call the user and check 1) buyer is available and picks up the call -simple solution (otp and done) 2)buyer is not availalable but picks up the call - a) reschedule it to some other day b) some one else is picking up in his place (some one else is there and there is friction as he is not there and he picks up the call ) 3)buyer is not available and does not pick up the call - (NDR has to be filed as the customer is unavailable )

Flow map: day1 user placed the order (protien powder · flipkart · 4 days delivery) → day2 and 3 shipped and dispatched → day 4 - morning out for delivery and the driver is assigned → driver will place the order based on the optimised route in a stack → - 2) buyer is not availalable but picks up the call → a) reschedule it to some other day · b) some one else is picking up in his place → OTP is shared / OTP is not shared (buyer is not picking up the call) but otp is shared later · c) leave it at the door/ security - shares OTP - 3) buyer is not available and does not pick up the call - cod and prepaid - (NDR has to be filed as the delivery is failed ) → recheduled to the next day - 1) buyer is available and picks up the call - simple solution (otp and done)-otp is shared - 4) cod (high AOV)- OTP +money - delivery guy doesn't know the location → ask on whatsapp/ call · mark as customer unavailable/ other issue

2.7 Return (day by day)

• day1: return because the protein powder is not the correct flavour → Schedule is written on the flipkart app → wrong flavor · wrong size (quanitity) → schedule for return with proof of image · damaged · not happy → ans a lot of question → schedule a return → return is scheuled on the expected date → delivery agent doesn't pick up the order: either you again raise a new ticket on flipkart and talk to all the chat bots / you just simply select on whatsapp to reschedule the next day. → the delivery agent picks up the order: delivery agent refuses to collect due to product mismatch or other reasons (wrong item, used item, etc) / delivery agent collects and takes photo and insepects · no return due to policy of "no return"

2.8 Voice agent and logistic agent contribution (purple and green cards, verbatim)

voice agent contribution: calling before hand for availability of the user sentimental analysis of the user (if he seems confused or not) having the data about his specific order understanding the state of the user. 1) if he/ alternative people are available, and leave at the door etc trying to decode what the user is saying in low connectivity and noise environment

(arrows:) context about the order is sent to the voice agent through logistic agent · optimizing the delivery guys's route based on the users current status if he is busy etc

logistic agents contribution - generating dynamic QR per order ( fraud and friction ) QR solves for fraud and friction Sending a trigger to the voice agent to make a call to the user qr solution- this has to be a picture sent to the customer or his proxy in the messages or whatsapp and not a link that needs to be clicked optimized dynamic route updating the Live location from Wa to delhivery's users history to permanently save the users location 1. leave at the door only - the picture of the parcel at the door (if the users tells to leave the parcel at the door when are we going to ask for alliterative numbers - whatsapp based question, or if there are multiple orders for the same customer , the bills will be combined into a same qr and is added into the metadata , the delivery guy will scan each barcode on the box and scan the qr simultaneously for every package

logistic agents for returns lets take a picture of the surrounding

Label on board: "Acts on its own inside limits somebody set. The person deals with whatever falls outside those limits." (the L3 definition)

Voice agent call flow (small mind map): call placed through voice agent - voice agent will call 2 times → - pick up → get the details from the user with 100 percent confidence → does the task as instructed · less than 80 percent confidence of the ai agent → delivery guy will call again to confirm - unhappy flow → 1st call → if doesn't pick up → call again - 2nd call → if doesn't pick up → delivery guy will call again manually · if he does not pick up

2.9 Agent - human interaction flow (framework, verbatim)

framework to think about agent human interaction

raj has ordered a pair of shoes which he expecting on day 3 and arjun is the delivery agent who will delivery him the show. this flow will help you understand how does the agent works

1. How the user initiates interaction.-first raj gets a whatsapp message sent by the logistic agent that if he is available or if there are any alternative people who will be available to pick up the order. This is where Raj then gives his mom's contact.

2. How the agent communicates with the user. - through voice agent - a call is placed (this called is shown in the called id as "delhivery agentt")

3. What information the delivery agent sees. - if raj is available or not with a confidence score. if the if the confidence score is less than 80%, then the delivery agent will have to call again and confirm.

4. When does the agent asks for approval. - a)To send the phone number of alternative people who will be available to collect the order or delivery Ask him to approve.

b)Approvals for Arjuns Gps (geofence)

c) to get the current data of the order that arjun has placed for a personalized voice call

5. When the agent acts independently.: when the confidence score is high from the voice agent that Raj or his mom is a) available to collect the order (after almost 2 calls are placed) b) not available to collect the order (after atmost 2 calls are placed

6. How the user handles exceptions or escalations. If the confidence score is <80% If, due to background noise or low connectivity, the voice agent is not able to determine the exact status, then the delivery agent Arjun does a manual call.

7. how does the agent interact with delivery agent If, let us say, The voice agent gives a low confidence score that Raj is available or not available. The dynamic route is not going to get changed, and Arjun, regardless, will still come to Raj and then place a manual voice call to confirm if he is available or not available, or leave at the door.

If there's a high confidence score (95% to 100%) that Raj is not available, then the logistic agent will dynamically reroute Arjun's route, will mark him as not available, and will move to the next delivery address after raj

Touchpoints (small map): delivery agent - H1 → point of touch - driver app/courier app → automatically and dynamically reroute his delivery route […] · end customer - (buyer/ alternative reciever) - H2 → point of touch - call/ whatsapp (voice rail) → first interaction is for […] triggered in the morning "[…] out for delivery" · ask: if you are not ava[ilable today, can you give us] alternative numbers w[hom we can contact for delivery] · voice agents calls the […] buyer first → if he does not pick up → alternative number provided in the first interaction → send the qr to either of these alternative contacts: mom · security · maid · neighbor · "send him a messages on his delivery app to call for low confidence of the voice agent (if the voice agent ain't that confident of status of the customer)" · "QR will sent as an ima[ge on whatsapp to the buyer/] alternative reciver" · "voice agent script (something like this): hey vardhan , your order for shoes is scheduled today and the agent will come to your door in an hour , are you available or any one else is available on behalf of you , should we leave it at the door?, reschedule it some other time you are available today or tomorrow ?" · "why voice agent will have a confidence level that then scores how confident it is that the end user is actually not avialable because we want to reduce postive and false negative[s]"

2.10 Stage flow charts (the white flowcharts on the right of the board)

flowchart TD
    subgraph S0[Stage 0 - Payment Setup]
        A0[Buyer selects Agentic COD at Flipkart checkout] --> A1[UPI Reserve Pay - SBMD - hold placed]
        A1 --> A2[Hold token bound to shipment AWB]
    end
    subgraph S1[Stage 1 - Alternate Capture]
        B0[Order marked out for delivery - morning WhatsApp] --> B1[WhatsApp: any alternate numbers?]
        B1 --> B2[Buyer registers alternates - mom, security, maid, neighbour]
    end
    subgraph S2[Stage 2 - Contact and QR Issuance]
        C0[Stack reaches T-5, five stops out] --> C1[Voice agent Call 1 to buyer]
        C1 --> C2{Buyer answers and confirms?}
        C2 -- Yes --> Q1[Logistic agent generates QR tagged to buyer]
        C2 -- No pickup, or unclear audio --> C3[Hold - stack keeps advancing]
        C3 --> C4[Stack reaches T-3, three stops out]
        C4 --> C5[Voice agent Call 2 - status check]
        C5 --> C6{Confirmed now?}
        C6 -- Yes --> Q1
        C6 -- No --> C7{Confidence buyer unavailable >= ~70-80%?}
        C7 -- Yes --> C8[Driver app nudges DA to call - route not paused]
        C8 --> C9{DA confirms recipient present?}
        C9 -- Yes --> Q1
        C7 -- No --> D1
        C9 -- No --> D1[Voice agent calls alternates from Stage 1 list]
        D1 --> D2{Alternate answers and accepts?}
        D2 -- Yes --> Q2[QR generated, tagged to that alternate]
        D2 -- No --> E[Escalate - see unhappy flow]
        Q1 --> QS[QR sent as WhatsApp photo]
        Q2 --> QS
    end
    subgraph S3[Stage 3 - Doorstep Handoff and Settlement]
        F0[DA enters 15m geofence - Agentic Safe Drop unlocks] --> F1[Scan order barcode]
        F1 --> F2[Scan recipient QR]
        F2 --> F3{Metadata matches authorised recipient?}
        F3 -- No --> F4[Reject and flag as fraud]
        F3 -- Yes --> F5[Camera PoD photo + GPS proximity confirmed]
        F5 --> F6[Pine Labs capture fires on UPI Reserve hold]
        F6 --> F7[Payment cleared - DA app flashes Safe to Leave]
        F7 --> F8[Order logged as delivered - buyer gets receipt push]
    end
    S0 --> S1 --> S2 --> S3

2.11 Field photos on the board

The board also holds screenshots and photos of a real delivery agent's phone: a Delhivery tracking page in WhatsApp's browser ("Arriving Today", Delhivery Maps), a delivery-rider rating screen ("Your Live Rating"), and the rider console showing "Contact consignee is pending" and "Geofence not [reached]" on a shipment card. Your label: "Mind map". These are the field evidence for the Evidence criterion; keep the originals in the folder.

Back to 00 README Index