A multi-product studio — SaaS, open-source, and commissioned work — built around senior engineering and opinionated design, with AI woven through the craft, not bolted on top.
We're not a generalist agency. We're a focused team that ships production AI, IoT, and web platforms — and pairs every engagement with the design intuition required to make it land.
Full service mapBuild with the smartest model in the room.
Production agents, RAG systems, prompt-cached pipelines, tool-use orchestration. We treat modern foundation models as a first-class platform — not a wrapper.
Hardware to dashboard, no missing layer.
Sensor meshes, edge intelligence, telemetry pipelines. Built on field-tested architectures — including a smart-metering platform recognised by the Ministry of Power.
Senior full-stack, end-to-end.
TypeScript-first apps, type-safe APIs, infra-as-code. We ship production from week one and refuse to leave a half-finished platform behind.
Opinions backed by craft.
Token-driven design systems, accessibility audits, and full UX teardowns of broken flows — the kind we publish as theses (see our IRCTC Tatkal teardown).
Boring infra, beautiful pipelines.
Multi-cloud architecture, CI/CD that actually catches things, observability that someone reads. Designed to be invisible — until you need it to save you.
Sensors, soil, sky — and the dashboard the field officer actually opens.
Low-power sensor meshes, low-bandwidth dashboards, and decision tooling for small-holder agriculture. Built on the same embedded and telemetry foundations as our smart-metering work.
A senior brain in the room when it matters.
For early-stage teams: architecture reviews, technical due diligence, and a senior engineering perspective on the decisions that compound — not a consultancy with a deck.
The best technology disappears. It doesn't announce itself with novelty or noise — it simply does the work, precisely as intended, leaving people the room to think.
A working studio doesn't get the luxury of vapourware. Every project below ran in production, served real users, and survived the messy bit after launch.
searoute-ts — the shortest realistic sea route between any two points on Earth. Eurostat network, canal and strait restrictions, vessel-draft gating, ETA from speed.
A keyboard-first Microsoft Teams client on the Graph API — AI summaries, voice rooms, disappearing DMs — with an MCP server so Claude and Cursor can use Teams as a tool.
A drag-and-drop email editor built once as web components, with wrappers for React, Vue, Angular and Svelte. MJML in and out, pre-flight linting, Figma import, MCP server.
Copy-paste UI for AI apps — tool-call cards, OAuth consent, scope inspectors, JSON-Schema tool forms. 38 components across React, Vue and Angular over one shared core.
promptbucket.co — an open library for collecting, versioning, and sharing prompts that actually survive contact with production. Built for the working AI engineer.
A drop-in prepaid meter that reads an existing analogue meter by its pulse LED and reports over MQTT. Published in IJCA; recognised by the Ministry of Power & Coal.
Listen, audit, draw the system.
A working week of stakeholder interviews, system maps, and assumptions held up to the light. We refuse to estimate before we understand.
Decide the bones before the pixels.
Architecture decided up front — model choice, data flow, deploy targets. Every diagram has a sibling document explaining the trade-offs.
Production from week one.
Vertical slice in front of users in the first sprint. Continuous delivery, tight feedback loops, telemetry from day one. No big-bang reveals.
You own it. We back it up.
Documentation pairs with the code. We train your team, sit in their pull requests, and stay on call until you decide we don't need to be.
A working studio writes — about the systems we admire, the ones we'd redesign, and the ones we wish someone would take a saw to. Long-form, slow-published, often uncomfortable.
Diagnosing Tatkal and designing what should replace it. Three concrete proposals — a two-track Tatkal lottery, a Booking Reputation Score, and an atomic booking pipeline that kills 'money debited, no ticket'.
The cheap way to build agents is to wire tools to a model and pray. The reliable way is a small constitution, hard tool boundaries, structured output, and a human in the loop wherever a refund is possible. Here is the pattern.
The most expensive bug we ever didn't ship wasn't in the code. It was in how we worked. A note on PR review queues, the cost of context-switching, and the small AI tool that bought a team back its senior engineering hours.
A small number of partnerships open at any time — usually one ambitious AI build, one platform engagement, and a UX teardown or two. Tell us where you are.