01
Enterprise AI & application architecture
The problemAI features get added to systems that were never designed for probabilistic components: no contracts, no failure boundaries, no way to explain a result.
What I doI design the contract first — what the model receives, what it may return, what happens when it returns something wrong — then place it in an architecture that already has observability, permissions and rollback. The AI is a component, not an excuse to skip design.
Evidence: Enterprise agentic engineering platform; DPDP compliance platform architecture.
02
Data intelligence & grounded AI
The problemOrganisations have documents, records and spreadsheets that disagree with each other, and answers generated over them are fluent and unfalsifiable.
What I doProvenance before generation. Every retrieved fact keeps its source location and extraction method; deterministic computation produces numbers; the model composes language around facts it can cite. If the system cannot cite, it says so.
Evidence: Compliance discovery and evidence mapping; news and content platform with publication gates.
03
Agentic workflows with bounded orchestration
The problemAutonomous agents are given broad goals, work confidently past their competence, and leave nobody able to say what changed or why.
What I doOne agent per self-contained task, frozen scope, explicit stop conditions, verification separated from implementation, and human approval at anything consequential. Parallelism only where tasks are genuinely independent.
Evidence: The platform this site was built with; documented mission contract and gate structure.
04
Private & local AI systems
The problemSensitive data cannot leave the premises, but the local stack is unmeasured — nobody knows what the model actually costs in memory or how it behaves under concurrency.
What I doQualify model and runtime combinations on the real hardware: throughput, effective context, memory residency, behaviour under concurrent load. Track residency and leases, isolate competing GPU workloads, and require a documented reason before anything is sent to a cloud model.
Evidence: Local-first AI platform on Apple Silicon; Qwen3.6 measured at roughly 51 tokens per second in the tested configuration.
05
Cloud, DevOps & infrastructure automation
The problemDelivery pipelines are slow, environments drift, and a recurring manual task quietly consumes days per release.
What I doEverything as code and reproducible: Terraform, Ansible, Packer, Bicep, containerised builds and pipelines that fail loudly. Measure the cycle before and after, because 'faster' is not a result until it is a number.
Evidence: 72h → 6h validation cycle; 9h → 5m permissions workflow; 5760h → 640h regression effort per release.
06
Quality engineering & enterprise troubleshooting
The problemA product crashes in a customer environment that cannot be reproduced, and the release clock is already running.
What I doReproduce from evidence — logs, dumps, environment diffs — then fix the cause rather than the symptom. Security defects get the same treatment: prove the injection path, close it, then test that it stays closed.
Evidence: Crash and hang diagnosis across APM, NSM and AutoSys; SQL injection, command injection and XSS fixes; regression automation.
07
Application modernisation & operational software
The problemLegacy enterprise software must keep running while its platform, architecture or language generation changes underneath it.
What I doIncremental, verifiable migration with the old and new paths running side by side long enough to compare behaviour. No big-bang rewrites without a rollback.
Evidence: Substantial contribution to a 32-bit to 64-bit enterprise product conversion; installer and pipeline modernisation at Progress.