Short essays and notes on software, quantum, AI, and the engineering practice behind them.
Rebuilding Lakera’s Gandalf prompt-injection game with a plain LLM to make failure modes concrete and highlight input/output controls.
When microservices help, when they hurt (always), and how to avoid a distributed monolith by defaulting to a modular monolith.
Why production outcomes lag demos, and how evaluation, data boundaries, and cost control determine adoption.
GenAI prototypes are fast, but production outcomes depend on evaluation, data boundaries, and cost control. I connect today’s enterprise adoption slowdown to older deep-tech patterns, including lessons from production video analytics and the AI infrastructure arms race.
Read the postMicroservices can help when domain boundaries are real and teams need independent deployment. More often they become a distributed monolith with expensive operations and coordination. This post lays out the warning signs and what I would do instead by default.
Read the postDeep tech makes it tempting to wait for clarity, but that usually produces demo-ware and long feedback loops. This post argues for shipping thin end-to-end slices early, and for including one real constraint in every increment so the system stays honest.
Read the postA practical follow-up on deploying LLM features: treat the model as untrusted, control both prompts and outputs, and assume cost is a product constraint. I map concrete mitigations to the OWASP GenAI Top 10 (2025) and discuss what to look for in an AI gateway.
Read the postI played Lakera’s Gandalf prompt-injection game and tried to recreate several levels with a plain LLM and a system prompt. The exercise makes prompt reinjection failures concrete, and highlights why you still need input/output controls beyond prompting.
Read the postAn overview of how LLMs handle prompts, why prompt injection works, and why multimodal systems widen the attack surface. I use Lakera’s Gandalf game as a concrete thread and end with the defensive patterns that actually matter.
Read the postTravel has one of the messiest product surfaces in tech: search, pricing, inventory, and customer service all collide. This post looks at where AI helps, where it breaks incentives, and what that means for platform architecture and user experience.
Read the postA reflection on abstraction in programming and mathematics, and on the tension between clean interfaces and performance realities. I connect this to AoS vs SoA, to optimizer placement, and to lessons learned building Qadence.
Read the postStarting a greenfield project is an iterative requirements exercise, not a one-shot design task. I share a practical approach to early PoCs, stakeholder alignment, and keeping ambition in check so the codebase stays easy to evolve.
Read the postReflections from IEEE Quantum Week, with a focus on the simulation workshop and the quantum software architecture track. My main takeaways: where performance tuning belongs, how usability interacts with quality attributes, and what the community is converging on.
Read the postNotes from ISC-HPC in Hamburg on where the HPC-quantum integration conversation is strong, and where I think it misses near-term algorithm work. I also share concerns around siloed European projects and premature standardization.
Read the post