Engineering
How we ship.
Real architecture decisions, real bugs we hit in production, real lessons from running a studio that ships parallel SaaS. No marketing recipes.
- aviation
DGCA compliance for flight data programs: what auditors actually look for
DGCA's FDR expectations are documented but spread across multiple CARs and advisory circulars. Here's the operational shape of what an audit actually checks, distilled from running an FDR analytics platform for the Indian market.
Indian carriers operating above defined thresholds are required to run flight data monitoring programs and produce evidence on demand. The DGCA Civil Aviation Requirements that govern this are real, public, and the standard reference. What's less well documented is the operational shape of an audit — what evidence the inspector actually asks for, what gaps come up most often, and what tooling materially reduces the friction. This post is the second.
Read more → - aviation
When does an MRO outgrow spreadsheets?
There's a specific moment when a maintenance shop crosses from 'Excel is fine' to 'Excel is dangerous.' Most shops don't notice it until an audit catches them. Here are the signs.
Aircraft maintenance shops live on spreadsheets longer than they should. The transition from 'Excel works' to 'Excel is a liability' isn't gradual — it's a discrete jump that happens when one of three things crosses a threshold: regulator count, fleet count, or workforce count. Once any of those crosses, the spreadsheet stops being a tool and starts being an audit risk. This is how to recognise that moment.
Read more → - markets
Why we don't let the LLM trade for you
Every retail investing app is racing to add an LLM 'agent' that places orders. We deliberately stopped one tier short. Here's the reasoning, and why we think the line we drew is the right one for everyday investors.
We built Sthira for layman retail investors in India. Adding an LLM that can place orders autonomously would be both technically straightforward and a category mistake. The mistake isn't in the technology — it's in the regulatory and behavioural model the product is trying to live inside. This is why we shipped automation up to the line of pre-authorized drift rebalancing and stopped there, and why we think you should be skeptical of any retail app that claims more.
Read more → - logistics
Why most TMS software fails operators
Transportation management systems get bought by procurement and used by dispatchers. Those are two different people with two different jobs. That's the whole problem.
Most TMS software is bought by people who don't use it and used by people who didn't choose it. The result is a category of products that demo well, fail in daily operations, and get worked around with spreadsheets. We've spent two years building against that pattern with Logic Nexus. Here's what we learned about why the existing options miss.
Read more → - aviation
What is ARINC 717? A practical guide for flight data engineers
If you've inherited an FDR analysis pipeline, or you're evaluating one, ARINC 717 is the wire format you have to understand. Here's what actually matters, without the spec PDF.
ARINC 717 is the binary wire format every commercial aircraft uses to write flight data. The full spec runs hundreds of pages; in practice you only need to understand half a dozen concepts to read what's coming off a recorder. This is what we wish someone had told us before we started building Aviation AI Pro.
Read more → - engineering
How we built FDR ingestion — and what shadow mode caught
We have a working Python parser. We had a finished C++ one sitting on the shelf. Here's how we wired them in parallel and what the comparison surfaced.
Aviation AI Pro has two ARINC 717 parsers — Python that ships and a C++ one that didn't compile until today. Wiring them side-by-side took fixing six long-standing bugs, rebuilding the upload pipeline through Supabase Storage, and adding a Celery shadow task. The shadow run on a real 18.8 MB .dlu file surfaced exactly the disagreement shadow mode exists for.
Read more → - engineering
Observability before customers
We wired Prometheus, Alertmanager, and Grafana into Aviation AI Pro before the first user signed up. Here's why, and the volume-mount bug that almost burned us.
Most pre-revenue startups treat observability as a Series-A problem. We treated it as a day-one problem because nothing we ship gets meaningful adoption without us first being confident it works. The wiring took roughly a day, fits in eight containers, and surfaced two bugs we wouldn't have found otherwise. It also nearly cost us all our dashboards on the second restart.
Read more → - engineering
Multi-tenant RLS at studio scale
Postgres row-level security is the simplest answer to tenant isolation. The simplicity ends the moment you have to migrate 46 tables under live load.
We run multiple Supabase projects, each with hundreds of tables and dozens of live tenants. Row-level security is how we keep tenant A from reading tenant B's data, but the path to fully-enforced RLS surfaced two problems we hadn't planned for: a 3-day migration drift between local and prod, and zombie Postgres connections that turned a routine ALTER TABLE into a 4-hour debugging session.
Read more →