The engineering laboratory

A standalone research track toward an energy monitoring platform for solar battery systems — monitoring, control and optimisation with lifespan prediction, built as a standard web/cloud application and designed to become the system of record for a site operator's installations.

• Live battery & solar telemetry, across sites
• Platform architecture designed — edge gateway, API, dashboard and data processing as one system
• Tiered retention & rollups — years of history at sensible resolution
• Two-signal health — heartbeat and data read separately, so a dead link is told apart from a dead meter
• Remote diagnosis — device logs buffered offline and replayed, so the cause arrives without a site visit
• From telemetry to action — operational state, safe remote action where the equipment supports it, verification and an evidence trail
• Lifecycle prediction & degradation analysis; cost & consumption tracking

View the dashboard prototype →

Next direction: the same platform applied to remote consumption monitoring & management — metering what remote appliances draw, switching them on/off, and invoicing per energy consumption.

DelfinFlow runs its own business on a defined, documented model — decisions recorded, handovers deterministic. The research direction is the software that runs that model: an auditable event log as the record, with a deterministic decision kernel for recurring questions.

• Guild — the operational runtime. The practice's delivery model in software: clients, projects and delivery as events and reactions, auditable at every step.
• Engram — the decision kernel. Decisions and memory in one auditable place, resolving recurring questions without an LLM call.

DelfinFlow's AI-assisted work runs on a convergent working profile of the practice: how it actually works, built from real sessions and recorded decisions, revised in place as it learns. The profile is carried into every session, so answers, automation and communication hold the same judgment the person behind the work would — better answer suggestions, processes automated the way the developer who built the product would automate them, emails composed in the practice's voice, and in clearly bounded cases, decisions taken on the user's behalf.

• Built from real work. Sessions and recorded decisions converge into the profile; it is revised in place, never appended.
• Partial autonomy. Suggestions, process automation, email and bounded decisions — the assistant acts on the user's behalf, within defined limits.

Measured, not asserted: over one 48-day period this profile kept99% of all input served from cache and cost under 5¢ per commit — every request, every model and every failed attempt included.

Pilot for your business: a convergent profile built around one real workflow — answers, writing and bounded decisions in your voice, in four weeks. It begins with a two-day kickoff — the structure set up and one flow working end to end, so you can judge by results before committing. Start a pilot →

* Pilot scope: one workflow, one profile. The two-day kickoff sets up the structure and one working flow; the four weeks converge the profile around that single workflow, not the whole operation.

The practice's own application-surface runtime: a generic UI shell that renders typed page definitions from the backend — operational interfaces, portals and dashboards without bespoke frontends. The companion monitoring tool tracks events and signals, not just log lines.

• ui-runtime-web — the declarative runtime. Backend-driven UI: typed page definitions interpreted by a generic React shell; the future surface for client portals and the energy dashboards.
• SurfLogs — signal monitoring. Developer-first monitoring of events and signals, with cascade detection and user-controlled issue grouping.