Insights
Thinking in public
We publish what we learn as we learn it. Not polished papers — working ideas, observations from the engineering floor, and the occasional strong opinion.
-
September 15, 2026
·SDLC
·LinkedIn
The Tech Team Was a Means to an End: Software. The End No Longer Needs the Means
Software has always carried two things bound together: the labour of building it and the judgment of what to build and how to structure it. For fifty years they were inseparable, so every methodology — Waterfall, Agile, sprints, standups, code review, handoffs — was really a workaround for how people think and coordinate. AI separates the two. The labour becomes automatable; the judgment does not. This argues that the lifecycle is redesigned, not accelerated: iteration moves from code to specification, where humans decide what should exist and how it should be built. As with cloud making compute abundant, the effect is more software, not less — the question shifts from 'what can we staff?' to 'what's worth doing?', and human judgment becomes the scarce constraint.
Read on LinkedIn →
-
September 15, 2026
·AI Agents
·LinkedIn
The Model Can Be Probabilistic. The System Shouldn't Be
As agents take on more of the build, the classical engineering disciplines matter more, not less. A single agent handed the whole codebase, every tool, and deploy authority is just the God Object returning in a new costume — probabilistic decisions with no clear line of responsibility make failures hard to diagnose. The argument here is to hold agents to output contracts rather than model personalities, give each one the narrowest access its job requires, and grow capability by composing specialised agents and validators rather than expanding prompts. The dividing line: prompts guide behaviour, software enforces invariants. Type constraints, schemas, approval gates, and evidence trails do the work that natural-language instructions can't reliably hold.
Read on LinkedIn →
-
September 1, 2026
·SDLC
·LinkedIn
AI Needs a New SDLC: A Complete Redesign, Not Just a Boost from AI Agents
Waterfall gave way to Agile, which shortened the feedback loop but left the underlying constraints intact: specialised teams, context lost at every handoff, rework discovered late, throughput capped by headcount. Dropping AI agents into that process makes the people faster without changing the shape of the work. This article argues for collapsing the lifecycle to two stages — stakeholders approve a prototype they have actually tested, then the system builds the production software — which moves the governing question from 'can we build this?' to 'should we build this?'
Read on LinkedIn →
-
September 1, 2026
·Enterprise
·LinkedIn
Enterprise Software Needs a New Production Model
Enterprise software rarely fails because the code was bad. It fails because the wrong thing got built — business context degrades through every handoff, so generating code faster only arrives at the wrong answer sooner. The argument here is to measure the human effort sitting between business intent and working software, then attack it in two places: discovery that settles ambiguity with stakeholders before any code exists, and a dark factory that builds, tests, and refines from the validated blueprint. The result is a third option between rigid SaaS and expensive integrators — software that fits the business rather than the reverse.
Read on LinkedIn →
-
July 23, 2026
·Systems
·LinkedIn
The Factory That Software Never Built
Manufacturing was systematized — the assembly line, lean production. Software never was: decades of better tools improved the individual stations while the factory floor itself was never designed, leaving delivery a craft process with high variance and unpredictable outcomes. This essay argues AI's real opportunity is not faster point tools but orchestrating the whole lifecycle as one engineered system, with deterministic outputs at every stage.
Read on LinkedIn →
-
July 17, 2026
·AI Engineering
·LinkedIn
Engineers Have Become Faster. But Software Delivery Hasn't
AI tools have made individual engineers dramatically faster — yet delivery timelines have barely moved. Coding is only 20–35% of the lifecycle, so even a tenfold coding speedup compresses the whole timeline by a few percent while the bottleneck migrates to discovery, review, testing, and integration. A pipeline is governed by its slowest stages, not its fastest; real acceleration means redesigning the whole lifecycle, not the busiest station.
Read on LinkedIn →
-
July 14, 2026
·Agile
·LinkedIn
Agile Solved the 1990s Software Problem. We Don't Have That Problem Anymore.
Agile fixed the defining problem of the 1990s: slow feedback loops on long waterfall projects. Today's problem is different — incomplete discovery at the start, which Agile treats as inevitable and pays for in rework, an estimated 30–50% of enterprise project effort. With AI, ideas can now be prototyped and stress-tested in hours, moving problem-solving to the stage where change is still cheap. The next two decades belong to a redesigned process, not a better Agile.
Read on LinkedIn →
-
July 6, 2026
·Software Delivery
·LinkedIn
Why Software Delivery is Broken
Roughly two-thirds of software projects end in partial or total failure, and large IT projects run 45% over budget while delivering half the value promised. This piece argues the causes are structural, not personal — fuzzy discovery, premature architecture decisions, fragmented tooling — and that an industry which treats unpredictability as inherent ends up building coping mechanisms instead of fixes. The way out is re-engineering delivery end-to-end.
Read on LinkedIn →
-
June 5, 2026
·Essay · 8 min read
The Electrification Moment
AI coding tools are electric motors bolted onto a steam-age process. The gains from electrification didn't come from the motors — they came from rebuilding the factory. The same is true now.
Read →
-
September 22, 2025
·Essay · 9 min read
What 'governance' means when AI writes your software
The CIO's governance concern about AI-built software is correct — but misdirected. The failure mode isn't AI-generated code. It's the absence of structured requirements management that most organizations already have.
Read →
-
December 15, 2024
·LLMs
·LinkedIn
Navigating the Future of B2B SaaS: Small vs. Large LLM Models
Large Language Models are more than just the latest tech trend — they're game-changers, especially in the B2B SaaS space. This post explores the real trade-offs between small, fine-tuned models and large general-purpose ones, covering latency, cost, accuracy, and when each approach wins. The answer isn't always 'bigger is better.'
Read on LinkedIn →
-
November 15, 2024
·LLMs
·LinkedIn
A Systems Approach to Using LLMs: Bridging the Gap Between Cool Demos and Real-World Deployments
LLMs dazzle in demos but stumble in production — hallucinations and unreliability stand between a cool prototype and a deployable product. This piece reframes the LLM as a "lossy medium," drawing on how the early internet engineered reliability atop unreliable transmission, and argues that a systems-level approach — guardrails, verification, and orchestration around the model — is what actually bridges the gap to dependable enterprise software.
Read on LinkedIn →
-
October 15, 2024
·GenAI
·LinkedIn
Why is Building Production-Ready Software with Gen-AI like GPT-4 So Challenging?
Everyone is excited about generative AI, yet few startups ship production-grade products on top of it. This article breaks down the three forces working against them — output reliability and hallucinations, stale knowledge from training cutoffs, and the complexity of wiring AI into existing enterprise systems — and explains why, for now, the most successful Gen-AI applications come from established players who can fold AI into what they already run.
Read on LinkedIn →
-
September 15, 2024
·LLMs
·LinkedIn
Rethinking LLMs: A Path to Accessible and Reliable GenAI
Building bespoke models from scratch is expensive and out of reach for most teams. This post makes the case for layering — enhancing existing LLMs with additional technology layers to sharply reduce hallucinations and operational cost rather than reinventing the model — opening a route to reliable, advanced AI that small and medium businesses can actually afford and trust.
Read on LinkedIn →
-
August 15, 2024
·GenAI
·LinkedIn
Embracing GenAI: A Strategic Imperative for CIOs
For CIOs, generative AI is no longer optional — it's a strategic lever for innovation and efficiency. This article lays out how to adopt it deliberately: identifying high-impact use cases, solving hard problems faster than traditional methods, and building scalable implementation plans backed by real change management, so GenAI's content generation and automation become a durable competitive advantage.
Read on LinkedIn →
-
July 15, 2024
·GenAI
·LinkedIn
From Silos to Synergy: How Generative AI is Redefining SaaS for the Agile Enterprise
Enterprises drown in fragmented SaaS tools that don't talk to each other. This piece argues for custom, modular micro-SaaS powered by generative AI — specialized applications that stitch into one integrated system delivering real-time insight and operational visibility — and positions AI-integrated micro-SaaS as the next evolution of enterprise software, replacing one-size-fits-all with something genuinely agile.
Read on LinkedIn →