Originally published at parvejshah.com/blog/engineering-precision-data-platforms-sft-rlhf by Parvej Shah.
The initial brief from GenMorphics AI Solutions sounded straightforward: build a dashboard to help their team coordinate work across a global network of domain experts. As we got into the details, "dashboard" turned out to be underselling it — this became a systems-architecture problem touching compliance, access control, and payroll as much as task management.
Skill-Scoped Task Routing
GenMorphics works with specialists across software engineering, mathematics, legal reasoning, and scientific writing. Routing doesn't try to auto-score someone's expertise — a specialist holds skills within categories, and eligibility for a task comes down to whether they hold the named skill in the required category.
The real design problem wasn't the matching logic — it was making the categories specific enough to mean something. "Knows Python" is close to useless as a routing signal. Splitting software engineering into language-specific tracks, and mathematics into calculus, linear algebra, and discrete math as separate categories rather than one "math" bucket, is what actually made routing reliable. The hard part of this system was taxonomy design, not algorithms.
Access Control as Data, Not Code
The platform serves people with fundamentally different access needs — domain experts who should see only their own tasks, project managers scoped to their own client portfolios, admins with full reach. Instead of hardcoding role checks scattered through the app, permissions live in the database: each role is a row with a set of resource.action.scope permissions — payment.read.own, nda.manage.all, and so on.
That decision paid off the first time GenMorphics needed a new role with a slightly different permission mix. It didn't need a deploy — it needed a database row. For an app that has to keep pace with how a growing team actually organizes itself, that flexibility mattered more than a marginally simpler hardcoded system would have.
Building an NDA Engine Instead of Buying One
Client task data — code, documents, domain material — can't reach a specialist until they've signed an NDA specific to that engagement. Rather than bolt on a third-party e-signature product, we built the NDA lifecycle inside the platform: a TipTap-based rich text editor with a custom node type for inserting recipient-specific variables into a legal template, template versioning so a signed document stays tied to the exact terms it was signed under, and a cryptographic hash of the final document for tamper-evidence.
Because NDA volume scales with contractor headcount, admins also needed to act on instances in bulk — voiding, expiring, counter-signing, reverting, or extending many at once rather than one dialog at a time. Building this in-house meant the NDA gate could be wired directly into task access instead of living as a disconnected compliance checkbox.
Payroll That Matches How the Org Actually Pays People
Payroll logic isn't one formula — it mirrors two different employment models in the same system. Operational roles (annotators, reviewers) are paid hourly against logged, effective time. Managerial roles are paid a fixed salary unless they have an hourly rate set, in which case hourly wins. Bonuses layer on top of either. None of this is generic timesheet-software logic — it's a direct encoding of how GenMorphics structures compensation across two different kinds of contributors.
What We Learned
The technical complexity here was never really about AI — it was a direct translation of how GenMorphics organizes people, compliance, and money. The skill taxonomy, the data-driven permission model, the in-house NDA engine, the dual payroll model — each one exists because encoding the org's actual structure mattered more than reaching for the nearest off-the-shelf pattern. A schema that's grown to 28 models across 61 migrations without a rewrite is really a record of that structure evolving over time, and the platform evolving with it.
Parvej Shah is a Lead Full-Stack Web Developer & Platform Architect based in Dhaka, Bangladesh. Explore full architecture case studies and production code at parvejshah.com.
Top comments (0)