Notes on deep agents, App Foundry, and the promptable platform.
How we think about building agents, tools, workflows and whole apps from prompts — and running them safely at enterprise scale.
GridHosting capacity, automated: a digital twin for DER interconnection
Topology in, feasibility out: an AC power-flow twin with IEEE 738 thermal limits and N-1 contingency turns interconnection studies from months into minutes.
GridCutting a feeder interconnection study from months to minutes
How a DISCOM used a power-flow digital twin to screen DER applications automatically — voltage, thermal and N-1 — and route only the hard cases to engineers.
Theft DetectionCatching electricity theft with behaviour, not accusations
A LightGBM detector that reads consumption behaviour — runs, entropy, lost weekly rhythm — and returns a ranked audit queue, gated on recall before it ships.
Theft DetectionTurning 42,000 smart meters into a ranked audit queue
Scoring a 42,000-customer meter panel end to end: features, an imbalance-tuned model, a recall gate, and a field-ready audit queue that respects inspection capacity.
Solar ForecastingForecasting solar you can bank on: quantiles, skill scores, and a release gate
Day-ahead solar generation as calibrated P10/P50/P90 quantiles, judged by skill score against persistence, and gated on calibration before any model ships.
Solar ForecastingA day-ahead view of solar output — with no data team
How an MSME manufacturer got calibrated day-ahead solar forecasts and tariff-aware load scheduling running on its own premises — no cloud, no data team.
Anomaly DetectionThe anomaly detector that doesn’t cry wolf
Robust statistics — MAD, a Hampel filter, prior-window z-scores — that catch real faults in generation and grid telemetry without drowning operators in false alarms.
Anomaly DetectionCatching a failing inverter before the revenue vanished
How robust anomaly detection separated a real inverter fault from ordinary cloud cover — and explained which it was — before the losses compounded.
Data HealingGarbage in, garbage out: healing sensor data before you forecast
A physics-informed, multi-tier repair pipeline that fills gaps by their length and cause — and proves its own accuracy against held-out synthetic gaps.
Data HealingRebuilding a month of missing meter data — and proving it
How a multi-stage reconstruction rebuilt weeks of dropped telemetry, using neighbouring sensors, and quantified its own accuracy so the repaired data could be trusted downstream.
Invoice ProcessingInvoice processing that survives real documents
Layout-aware extraction across many document providers, schema-driven fields, validation rules, and exceptions that route themselves to a human — not a demo that only reads clean PDFs.
Invoice ProcessingThousands of invoices a run, with exceptions that route themselves
How a bulk invoice-validation pipeline mapped thousands of rows to a canonical schema, applied a dozen rules, and sent only the exceptions to a human — with an audit trail behind every decision.
Claim ProcessingClaims adjudication as a governed agent workflow
Source-grounded extraction, a ‘no citation, no claim’ discipline, and human approval wired in as a step — adjudication you can defend, not a black box that pays.
Claim ProcessingFrom claim inbox to decision, with a human on the line that matters
How a claims team automated intake and first-pass adjudication with cited extraction and confidence-based routing, keeping humans on the high-value and ambiguous cases only.
Licence ForecastingForecasting the renewal wave — then staffing for it
Two-mode renewal-volume forecasting (portfolio expiry projection + trend/seasonal history) feeding Erlang-C desk sizing: offered load, p90 wait, and the reviewers it takes.
Licence ForecastingStaffing a licensing backlog before it becomes a breach
How a compliance operation turned an expected renewal wave into a concrete headcount plan — p90 wait, SLA risk, and the seats needed to stay inside target.
App FoundryHow App Foundry builds a running app from one prompt
PRD → blueprint → scaffold → generate → verify → deploy — and why the output is an app, not a chatflow.
Deep AgentsThe deep-agent harness: ReAct, sub-agents, memory & HITL
A durable loop that plans, spawns specialists, uses MCP tools, and self-verifies — the substrate behind every build.
WorkflowsWhy a flow on a canvas isn’t a product — and what is
Visual builders stop at the graph. We compile the workflow into a governed, shipping app.
RAG & IDPGrounding agents in messy enterprise documents
MinerU, Docling and OCR turn real PDFs into structured, citable context you can build on.
ConnectorsEvery integration as a governed MCP tool
Provisioning connectors through a unified gateway — with credentials behind an HITL boundary.
GovernanceAutonomous builds, full accountability
HITL gates, RBAC, tenant isolation and a unified run ledger — autonomy you can audit.