Wrestling with Integration: Why My Project Feels Like a PhD (and Why That’s Okay)
I’ve been thinking a lot about why The Planner’s Assistant sometimes feels less like a project and more like a hydra — every time I get one prototype working, two new problems grow back in its place. Parsing local plan policies? Solved. GIS constraints? Solved. Officer-style reasoning? Solved. But the moment I try to integrate them, the whole thing feels sprawling and unmanageable.
Here’s the truth I’ve had to admit: the integration itself is the research problem.
From slices to systems
Over the past months I’ve built what I’d call “minimum viable slices”: a pipeline for extracting structured policy metadata, a GIS layer manager, a hybrid RAG retriever, an officer report generator. Each one stands up on its own. The difficulty begins when you try to stitch them into a whole.
It’s tempting to treat that stitching as “glue code,” but it’s actually the frontier: how do you make probabilistic LLM reasoning interlock with statutory policy anchors, or have spatial constraints flow into narrative officer reports without breaking explainability? These aren’t just engineering annoyances — they’re deep questions about how we formalise judgement in digital systems.
Integration as frontier science
Most AI research focuses on narrow capabilities: better models, bigger datasets, sharper benchmarks. Planning exposes a different challenge: multi-modal, multi-domain integration under accountability constraints. It’s not enough to parse text or query a map — the system has to braid them into a judgement that can withstand scrutiny.
That’s why it feels PhD-level: because it is. This is not “feature creep” or a lack of discipline; it’s what happens when you deliberately pick a domain (planning) where the real bottleneck is institutional rationality itself.
Reframing the work
So I’m choosing to see it less as “insanity” and more as excitement:
- Each integration point is a new research contribution.
- Each joint I make less leaky is more valuable than any shiny prototype.
- The sprawl is not a sign of failure — it’s a sign I’ve landed in the exact spot where no one else has charted the territory.
It also helps to call this a lab rather than a project. A project implies finishability; a lab implies exploration, coexisting prototypes, and gradual synthesis.
The stakes
If integration succeeds here, it won’t just mean a better planning tool. It could offer a model for other domains where AI has to grapple with law, policy, and discretion. It’s a testbed for building digital systems that are not just capable, but explainable and accountable.
And that’s worth the frustration.
What’s next (for me, and for the lab)
- Stabilising the joints: focus not on new prototypes, but on making two or three existing ones work together cleanly (policy metadata → RAG → officer narrative).
- Audit trails by default: bake in provenance tracking now, so integration doesn’t collapse under opacity later.
- Golden set testing: build a benchmark suite of nasty PDFs and policy structures so I can measure when integration actually improves.
- Narrative logging: keep writing posts like this, because externalising the struggle is part of how I keep the excitement alive and avoid getting lost in the sprawl.
👉 So yes, it feels like a PhD. But that’s because it is frontier integration work — and planning is the crucible I’ve chosen to do it in. The hydra is an orchestra in disguise.