SDSL
A case-management platform for a large private archive: the assets it tracks, the people and companies around them, and the documents, emails and recordings behind each one, all linked in one data model.
- Role
- Full-stack
- Client
- Novenext
- Year
- 2024 — 2025
- Stack
-
- Application
- Next.js, TypeScript
- Data
- PostgreSQL, Prisma, AWS RDS, AWS S3
- Infrastructure
- AWS EC2, nginx, PM2, Route 53, Let's Encrypt
- Services
- AssemblyAI, OpenAI, Dropbox API, Gmail API
- Availability
- Details under NDA. Described here without the client's commercial specifics, their customers, or the internals of their systems.
The problem
A large private archive accumulates 2 things in parallel: the assets it tracks, and the paper trail that explains them. Each asset has a history: where it came from, who has held it, what it’s worth. Around it are people and companies in various roles. Underneath sits the evidence: invoices, email threads, recorded conversations.
All of it was held in places that didn’t know about each other: files in cloud storage, correspondence in mailboxes, recordings scattered about. And the part that mattered most, that this asset relates to that company through that person, lived only in the heads of the people who had done the work.
What they needed wasn’t a document store. It was for the relationships to be first-class, so that a case pulls together everything connected to it, and an asset carries its own history.
What I built
Everything starts from a case. A case has a reference, a label, tasks, the people involved, the events it covers, and every document, email and recording that bears on it. Open one and all of that is on the page.
Events are the other half. An event is something that happened, on a date and at a place, linked to the people, assets and files it touched. Open an asset and its events read as its history, in order.
Evidence comes in from where it already lives. Mailboxes are connected and imported by label in the background, with a live view of every import that can be paused, resumed, retried or cancelled. Documents are uploaded or arrive as attachments. Recordings are picked from a linked cloud storage account, transcribed, and summarised by a language model, so a long conversation becomes something you can skim and search like everything else.
People, companies, vendors and places are records of their own, linked to everything they’re part of. The main kinds of record each have their own search, email has an advanced one on top, and any asset can be exported as a PDF record sheet.
Decisions that mattered
Relationships first. The quick way to build a tool like this is a set of separate lists, one per kind of record, each with its own search. That answers “find me this document”. It can’t answer the questions the archive actually asks, which are “what connects this person to this asset”, because a list per kind of record leaves a person to hold the connection in their head. So the links themselves are records, between cases and the people, events and evidence in them, and between events and whatever they touched. It costs a lot of join tables. The alternative leaves the same complexity with the user, where it’s invisible and never goes away.
Imports are jobs, not requests. A mailbox import is long and fallible. It can pull thousands of messages, fail halfway, or need stopping. So each import is a record a user can watch, pause, resume, retry or cancel, carrying its own status and history, and one that stops making progress is caught and paused rather than left hanging. The simpler version is a request that loops until it finishes, which has nowhere to put “partway through, and 3 messages failed to parse”. Recordings are safe to process twice for a similar reason: each one keeps its storage file’s own ID in a unique column, so the same file can’t come in as a second copy.
Operating it
It ran on AWS, and that side was mine as well as the code: the infrastructure, the database and the deploys.
Outcome
I was the only developer, from an empty repository to the system that ran in production.
The domain model turned out to be the heart of it. Almost every question the archive had to answer came down to whether the schema made a relationship explicit or left it implied. The links between records, the event model and the case as a first-class object took the longest to build, and they’re what made the rest of it simple.