Bullet Summarization - Postmark Challenge

This is a submission for the Postmark Challenge: Inbox Innovators.

What I Built

Bullet Summarizer is a summarization-as-a-service solution that acts based on email forwarding. The main goal is to drastically reduce the time needed to catch up on lengthy email threads, especially when multiple participants are involved.

The application can be directly integrated into any existing system, leveraging AI capabilities (any LLM can be integrated - ChatGPT has been used for the demo due to ease of implementation). All the data collected can also be stored into a client's databse for future reporting, analysis and tracking.

The core use case consists of saving time and reducing decision fatigue through a clean and structured view on relevant email threads.

Demo

The source code can be found here.

Any user can forward a thread of emails to the following address: [e-mail]. Due to domain issues and outbound rules, some users may not receive a reply with the summarization. In this case, the safest way to demonstrate the functionalities would be to replicate the runtime into your own environment. In order to replcate locally, fill in the appsettings.json file with the Postmark and ChatGPT Keys, together with the connection string.

In order to demonstrate the app's functionalities, I have opted for two different scenarios. Each scenario was based on fictive personas, so all names and examples should be taken with a grain of salt. The initial thread will be attached as a spoiler, in order to minimize the content.

In case of failure when sending the results to the user's email, I have created a simple Metabase dashboard that consists of all the tables from the database. All new entries should be automatically added to the dashboard: Metabase Dashboard.

For hosting I used a small Hetzner instance, directly exposing the Metabase UI and the API.

Scenario 1. App Launch - Team discussing bug fixes and timelines

The initial thread of emails:

scenario-1

============ Forwarded message ============
From: AlinPisica WebsiteContact <[e-mail]>
To: "testmail janedoe"<[e-mail]>
Cc: "Alin Pisica"<[e-mail]>
Date: Sun, 01 Jun 2025 22:33:22 +0300
Subject: Re: Urgent: FitFit Launch Status – Bugs Still in QA?
============ Forwarded message ============


Will do, Jane.
I’ll prepare a fallback launch plan (Aug 21) with revised Marketing and QA checkpoints just in case. Will share with you tomorrow morning.
Also, I suggest we lock down any non-critical merges starting Monday. Alin, let’s finalize that cutoff list on our 10AM sync tomorrow.
We’ll make this launch happen 💪

Cheers,
 —
 Mike Rays
 Chief Technology Officer
 📞 [phone]
 ✉️


În dum., 1 iun. 2025 la 22:33, testmail janedoe < mailto:[e-mail] > a scris:

Thanks both, this gives me more confidence.
Mike – let’s get a contingency plan in place anyway. Can you prep a quick internal doc outlining what the “Plan B” would look like if we needed to delay by 5 days?
Alin – great work jumping on these issues. Please send me a short daily update until the weekend. Doesn’t need to be long – just a few bullet points so I can keep stakeholders in the loop.

Thanks,
 —
 Jane Doe
 Chief Executive Officer
 📞 [phone]
 ✉️


On Sun, Jun 1, 2025 at 10:32 PM Alin Pisica < mailto:[e-mail] > wrote:

Hi Mike, Jane,
Quick update on my side:
✅ iOS crash: root cause found (Bluetooth permission regression in the new onboarding flow). Fix is in progress, should be in QA before lunch.
🔄 Sync issue: looks like the wearable module sends a malformed timestamp when the phone app is in background mode. I'm working on a patch that auto-corrects it on the mobile side. ETA: end of day.
If we don’t find anything new, both of these will be resolved this week.
Let me know if anything else pops up.

Best,
—
Alin Pisica
Software Developer
📞 [phone]
✉️


---- On Sun, 01 Jun 2025 22:31:55 +0300 AlinPisica WebsiteContact < mailto:[e-mail] > wrote ---


Hi Jane,

I understand the concern. I’ve already gone through the high-priority issues. Out of the 7:

3 are already in review, fix is done.
2 are visual bugs on smaller Android screens – not launch blockers in my opinion.
1 involves syncing between wearable and mobile – that one needs deeper investigation.
1 crash reported on some iOS builds – flagged by QA yesterday.

Alin is working on the iOS crash right now. I’ll let him chime in on the sync issue.
We should be fine for 16th if we stay focused. I don’t recommend postponing yet.

Cheers,
 —
 Mike Rays
 Chief Technology Officer
 📞 [phone]
 ✉️


În dum., 1 iun. 2025 la 22:31, testmail janedoe < mailto:[e-mail] > a scris:

Hi Mike, Alin,

I’ve just reviewed the QA dashboard this morning and noticed that we still have 7 open bugs marked as “high priority”. With only 15 days until launch, I’m getting worried.

Can we please regroup and assess:

- Which of these are actual blockers?
- What’s realistically going to be fixed before the 16th?
- Do we need to consider pushing back the launch?

Let’s sync before EOD. I’d appreciate a written update before 2PM so I can align with Marketing.

Thanks,
 —
 Jane Doe
 Chief Executive Officer
 📞 [phone]
 ✉️

Result:

Result as text:

scenario1text

High-Level Summary:
    The team is working on critical bugs for the upcoming launch and preparing contingency plans. There are discussions on bug fixes, release timelines, and communication updates.

Tone:
    The email exchanges are focused and collaborative, addressing urgent launch preparations and troubleshooting.

Participants
    Alin Pisica ([e-mail]) - Role: Software Developer
    Mike Rays ([e-mail]) - Role: Chief Technology Officer
    Jane Doe ([e-mail]) - Role: Chief Executive Officer

Action Items
    Prepare fallback launch plan: Develop a revised launch plan with checkpoints and an August 21 deadline.
    Assigned to: Mike Rays, Due: 2025-06-07

    Finalize cutoff list for non-critical merges: Determine which code merges can be blocked starting Monday.
    Assigned to: Alin Pisica, Due: 2025-06-07

    Send daily update on bug fixes: Alin to provide brief daily status updates until the weekend.
    Assigned to: Alin Pisica, Due: 2025-06-07

    Create contingency plan document: Outline Plan B for a potential 5-day delay in the launch.
    Assigned to: Mike Rays, Due: 2025-06-07

Key Dates and Deadlines
    Fallback launch plan readiness: 2025-06-07

    Final cutoff list for code merges: 2025-06-07

    Daily bug update reports: 2025-06-07

    Contingency plan document completion: 2025-06-07

Bullet Points
    Development team is fixing the iOS crash caused by Bluetooth permission regression and working on a patch for wearable synchronization issues. (Related to: Bug fixes and patches in progress.)

    Seven high-priority bugs are still open; some are review-ready, others require further investigation. (Related to: Current bug status and issue prioritization.)

    Management is considering a last-minute contingency plan for delay but currently aims to proceed as scheduled. (Related to: Risk management and launch planning.)

    Team members are providing daily updates to keep stakeholders informed about progress and issues. (Related to: Stakeholder communication.)

Scenario 2. Mortgage approval - Renegotiation of a property

The initial thread of emails:

scenario-2

============ Forwarded message ============
From: Alin Pisica <[e-mail]>
To: "testmail janedoe"<[e-mail]>
Cc: "AlinPisica WebsiteContact"<[e-mail]>
Date: Sun, 01 Jun 2025 22:39:34 +0300
Subject: Re: Mortgage Credit Approved – Next Steps for Purchase of Property on Linden Street
============ Forwarded message ============

Hi Jane, Mike,

Thank you both – that’s very generous and reasonable. I’m comfortable covering the additional €2,000 from savings and happy we found a middle ground.
Mike – please move forward with the updated paperwork for €153,000 and send over the revised credit agreement when ready.
Jane – I’ll ask my notary to propose a few slots between 6–9 August.
Excited to get this done!

Cheers,
—
Alin Pisica
📞 [phone]
✉️

---- On Sun, 01 Jun 2025 22:38:56 +0300 testmail janedoe < mailto:[e-mail] > wrote ---


Hi Mike, Alin,

Thanks for the transparency.
Alin – I understand this has been a process for you, and I don’t want to block your efforts. I’m willing to accept €153,000 and move forward, if we can schedule the notary before August 10th. That way I can align with my own moving timeline.
Let’s finalize this – I’d be happy to sell to you.

Best,
 —
 Jane Popescu
 Property Owner
 📞 [phone]
 ✉️


On Sun, Jun 1, 2025 at 10:38 PM AlinPisica WebsiteContact < mailto:[e-mail] > wrote:

Hi Alin, Jane,

Alin – after a quick internal review, we can approve an additional €5,000, bringing your credit ceiling to €153,000, under the same conditions (interest rate and term). However, you would need to cover the remaining €2,000 difference through your own contribution.

Jane – would you consider splitting the difference with Alin to help close the sale? Perhaps meeting in the middle could make this work for both sides.

Let me know your thoughts so we can move quickly.

Regards,
 —
 Mike Rays
 Loan Officer – Residential Mortgages
 FirstBank Europe
 📞 [phone]
 ✉️


În dum., 1 iun. 2025 la 22:38, Alin Pisica < mailto:[e-mail] > a scris:

Hi Jane, Mike,

Jane – I respect your position, but this comes as a surprise. We agreed on €148,000 and I’ve already paid for valuation, legal checks, and initiated paperwork with the bank based on that amount.

Mike – Is there any way the bank can increase the approved amount to match the new price? Otherwise, I’d be forced to withdraw.

I still really want the house, but I’m already stretched with my budget.

Thanks,
—
Alin Pisica
📞 [phone]
✉️

---- On Sun, 01 Jun 2025 22:37:30 +0300 AlinPisica WebsiteContact < mailto:[e-mail] > wrote ---

Dear Jane, Alin,

I’m pleased to inform you that the credit application for Mr. Alin Pisica has been officially approved. The credit covers the initially agreed price of €148,000 for the property located at 24 Linden Street.

Next steps:

We can begin preparing the pre-contract.
The notary appointment can be set once Jane confirms availability.
Please confirm if there are any last-minute changes to the agreed terms.

Looking forward to your confirmation to proceed.

Kind regards,
 —
 Mike Rays
 Loan Officer – Residential Mortgages
 FirstBank Europe
 📞 [phone]
 ✉️

Result:

Result as text:

scenario2text

High-Level Summary:
    Alin agrees to proceed with a property purchase at the approved mortgage amount of €153,000 after some negotiation. The next step is scheduling a notary appointment between August 6-9.

Tone:
    Informal and cooperative tone, with Alin expressing readiness to move forward and finalize the sale.

Participants
    Alin Pisica ([e-mail]) - Role: Buyer
    Jane Popescu () - Role: Seller
    Mike Rays ([e-mail]) - Role: Loan Officer

Action Items
    Send updated credit agreement: Mike to prepare and send the revised credit documents reflecting €153,000 limit.
    Assigned to: Mike Rays, Due: 2025-06-10

    Schedule notary appointment: Jane to propose options for notary appointments between August 6-9.
    Assigned to: Jane Popescu, Due: 2025-07-31

Key Dates and Deadlines
    Notary appointment between: 2025-08-06 to 2025-08-09

Bullet Points
    Alin agrees to move forward with the purchase at €153,000 mortgage limit. (Related to: Alin Pisica)

    The bank approved the mortgage for €148,000 initially, then increased to €153,000 after negotiation. (Related to: Mortgage approval)

Code Repository

The code is structured using a 3 layered architecture, split into three projects: API (that acts as the main app - exposing the API and using the background workers), Core (that hosts all the domain logic, models, types) and Infrastructure (the communications, integrations and processes).

The database migrations are created using DbUp, opting for a simple schema regarding the storage of all the results. No foreign keys or indexes were used in order to keep the system simple for demo purposes, easy to reproduce/run locally and avoid any possible issue regarding database optimizations. All the database migrations are located under Infrastructure/Repository/Database/Scripts. The queries and database access is simplified by using Dapper.

In order to host it on linux as a service, utilities are included into the utils folder (bash scripts for starting, stoping and checking the service's status, together with the service template to run under systemctl).

How I Built It

Tech Stack

Tech Stack used:

• .NET 8 for the API, processing and orchestration
• PostgreSQL for storing the inbound emails and summarization results (action items, bullet points, etc.)
• Structured Chat GPT for the NLP management, understanding and extracting information
• Metabase for dashboard visualisation

Architecture

The architecture consists of a .NET app that has the following three functionalities:

• a publicly exposed API endpoint that is used as a webhook for Postmark's connectivity
• a background worker that retrieves unprocessed emails from the database and stores them into an in-memory queue
• a background worker that retrieves data from the in-memory queue, computes the summarization, saves the results into database and sends the email to the user

Structured GPT Response

In order to easily obtain a consistent response, I have leveraged the functionality of structured answering of ChatGPT. In this way, every response is returned using a stable data structure that can be easily interpreted and mapped to our needs.

The response schema can be found here

Upcoming features

Since I loved working with Postmark, I am thinking of extending the current product with the following functionality:

• update the reasoning service to easily integrate with multiple LLM providers
• integrate directly with Google Sheets and Google Tasks to directly upload and manage data based on email interactions
• create multiple forwarding rules and differential results for all participants (each thread participant will receive its own copy of the summarization but only with its personal interests)
• add follow up functionalities - the system should follow up on previous summarizations to make sure that all the action items have been completed

Comments

[Nikoloz Turazashvili (@axrisi)]

Hey! Love the implementation; it truly shows skill.

I recently launched the Axrisi browser extension that transforms text on any website in the same way as in your submission, but I ran into an issue with Gmail specifically. :)

It already has Gemini inside the client, and practically neither my extension nor your app is really useful.

sorry if it might look like I'm judging. Commenting just to launch conversation, maybe you have some ideas to share as well. :)

[Alin Pisica]

Hey there and thanks a lot for the involvement! <3

I agree that lots of features are already included by some providers and your plain example with Gemini is a very strong and valid argument. However, I've lately seen a growing trend where multiple companies are going into the on-premise world, due to privacy issues and requirements. Lot of companies tend to store their own information in order to make sure that they are fully protected regarding their data distribution, so having a product that can run in house without giving crucial details to external providers can prove to be useful sometimes (as mentioned, the GPT can be replaced with any existing on-prem or cloud provided AI Reasoning service, and all the data can be stored in customer's own database).

You don't usually run against Google, Meta or the big ones, because there is a big chance you might lose - especially in the beginning. But you can play along them and target features that they do not provide. The summarization is an example, but the features can be extended in such a way that an app / a user could benefit both from the already included features of already existing, stable apps, but also from upcoming ideas and implementations. For example the app can target custom summarization, in specific formats, or tailored integrations that can create real value for specific user needs.

Just because something similar exists doesn't make for useless products. Lamborghini started to develop cars alongside Ferrari, both made fast and good looking cars but targeted different audiences. Xbox vs PlayStation is an ongoing battle and we don't see any of them going out of service soon. Competition breeds creativity.

I took a look at your extensions and mate, it is awesome. Are there similar things on the market? Yes. But I see features into your extensions that others don't bring, and for this I consider it special. Your work is solid and for sure you invested a lot of thought and time into it. Maintain your motivation high because great things are coming!

[Nikoloz Turazashvili (@axrisi)]

Thanks a lot for the kind words :)

That's a valid point; a tool like this might be useful for enterprises due to privacy concerns :) Furthermore, email communication with internal LLMs might be alongside the internal LLM chats.