Pillar / what an engagement is
Most AI work fails at the seam between the model and the system it is supposed to live in. We build the system first and put the model where it earns its place. Eight platforms are in production right now, and Ommune still operates every one.
A feature is a model call attached to a screen. A system is the thing that has to be right when nobody is watching: the data model that can represent the answer, the deterministic path that handles the ninety percent of cases a model shouldn't touch, the confidence boundary that decides when to escalate, and the fallback that keeps working when the provider returns a 500.
In Arthastra a misclassified bank transaction posts money to the wrong ledger and someone's books stop balancing. So the model never posts anything. It proposes, a rules engine and a human confirm, and every failure path returns the rules-based answer rather than nothing. That constraint — what happens when the model is wrong — is the actual design work.
Four things, and they usually arrive together in one system.
Turning unstructured input — a bank narration, a statement PDF, a CRM record, a free-text message — into a typed decision the rest of the system can act on. Tiered by cost: cheap deterministic matching first, learned patterns next, a model only on genuine ambiguity.
Intent classification, slot extraction, and a constrained executor that can only call a fixed set of real API operations. Agents that act on a system rather than talk about it, with the blast radius defined in code rather than in a prompt.
Multi-tenant architecture, double-entry integrity, idempotent posting, migration trees, adapter isolation so client-specific logic never leaks into core. The AI layer is worth nothing on top of a data model that can't hold the answer.
Routing between free, cheap and capable models by confidence and language; capping context deliberately; measuring what each call actually costs. In Arthastra the paid model is a fallback that fires on a small fraction of calls, at roughly $0.0001 each.
Named plainly, because most of this gets sold under vaguer words.
Six of the eight, written up in full: problem, architecture, AI components, what it runs today, outcome.
Two to three a year, which is the honest number for a team that also operates everything it ships.
Budget, timeline, and whether IP or equity is on the table. If the shape is wrong we say so in the first conversation rather than the third.
The data model, the deterministic path, the decision boundary, and the failure behaviour — written down before any model is chosen.
We build it. The engineers who scoped it are the engineers who ship it — no account manager in between, no junior tier the work gets handed down to.
Live, then kept live — reconciliation runs, audited data cleanups, model cost tuning, and the long tail of things production finds.
Model-agnostic by design — Claude as the default stack, OpenAI where the ecosystem fits, Sarvam for Indian languages, Cerebras and Groq where latency is the product. And the 90% of the work that is not model choice.
Consent infrastructure, automated data principal rights, retention pipelines that actually delete, and breach workflow — built into the product. Enforcement begins 13 November 2026.
There is no pricing page. Every one of these systems was scoped to a specific operational problem, and the number that follows from that has nothing to do with a tier on a page. Tell us the budget range you have and we’ll tell you whether it builds the thing you described.
Capacity: 2–3 engagements per year
Start a conversation →Budget range, timeline, and whether IP or equity is negotiable. Three questions, asked first because they are the ones that decide.