Intelligence · Platform

JJ Atlas

The agent layer: retrieval the database scopes to one tenant, model routing by task class, golden evals in CI, and a spend ceiling per agent per day.

Scope an Atlas engagement

Intelligence Intelligence

What it is

JJ Atlas is the agent layer: retrieval the database scopes to one tenant, a router that sends each task class to the model it actually needs, golden evals that run before a prompt change merges, and a spend ceiling per agent per day. Every answer cites the passage behind it or escalates to a person. That is the whole difference between a demo and something you put in front of a customer.

Reference architecture

What the platform is actually made of. You can read this before you sign anything, and you own it afterwards.

01Model routerTask class picks the model, never preference: reasoning to Claude Opus 5, high-volume classification to Gemini Flash, long-context synthesis to Gemini. Model ids live in one file, so a swap is one diff.
02Tool sets and autonomyAn agent that may not send has no send tool. Four levels from disabled to autonomous, checked before any external side effect, because a prompt is not a security boundary.
03RetrievalChunks embedded with text-embedding-3-large and stored as halfvec(3072) under an HNSW cosine index. Changing the embedding model is a full re-index, never a mixed one.
04Tenant isolationThe match function runs SECURITY INVOKER, so RLS removes other tenants' rows before ranking. Visibility is never a WHERE clause the application has to remember to add.
05Eval harnessAt least three golden cases per agent, run in CI against recorded transcripts. A prompt version that regresses a case does not merge.
06Cost and auditEvery run records model, prompt version, tokens and cost against a daily ceiling. Past the ceiling dispatch refuses, rather than spending quietly until the invoice arrives.

What you get

An agent registry: one row per agent with a versioned prompt, a typed tool set and a daily spend ceiling.
Your documents chunked, embedded and re-embedded on publish, with visibility derived at embed time rather than asked for at query time.
A retrieval isolation test in your suite: two tenants, identical content, one of them retrieves nothing.
Golden eval cases per agent running in CI, so a prompt edit that breaks a good answer is caught before it ships.
Answers that cite the passage they came from, and escalate to a named person below the confidence threshold.
Cost reporting by agent, model and day, against the ceiling each one runs under.

What it costs

Published bands with their assumptions. A platform page without one is a draft and does not publish.

Published price bands and the assumptions behind them
Band	From	What it buys	Typical duration	Assumes
Pilot	₹4,50,000	One agent in production against one corpus: retrieval, citations, evals, and human approval on every outbound message.	3–5 weeks	Up to 5,000 documents in text formats — scanned paper is an OCR engagement first One language Model provider accounts in your name, usage billed to you at cost Approve-before-send for the first month, reviewed after
Agent layer	₹14,00,000	Three to five agents across a workflow: routing, write-back to your systems, approval queues, dashboards, and the eval suite behind them.	8–12 weeks	Source systems expose a documented API, or a database we can read One corpus owner who can rule on what a correct answer looks like Evals authored with your team — we cannot write the golden cases alone Voice and messaging channels are a JJ Tech Connect engagement
Evals and cost governance	₹1,80,000 per month	Prompt versions, eval runs, routing review, and a monthly cost report as model prices and capabilities move.	6-month minimum	An Atlas layer we built, or one we have audited first Model price changes are passed through, not absorbed Two prompt revisions per agent per month Thirty days' notice to stop after the minimum term

Where it has run

48 staff

Three days of stock guesswork, gone

A two-floor clothing retailer replaced a Monday morning reorder meeting with a forecast read from its own till data.

31% target reduction in stock-outs across 400 SKUs

150 staff

Thirty-eight minutes to clear an exception, down to four

A freight operator's exception desk stopped being four people reading four screens.

4 min target median time to clear an exception, against 38 observed

55 staff

Every recommendation carrying its reasoning

An advice firm made the suitability basis a by-product of giving advice rather than a document written afterwards.

100% of recommendations carrying a cited suitability basis

35 staff

Proposals that stopped eating the fee-earners' evenings

A consultancy assembled proposals from its own past work instead of from last week's near-miss document.

6-9 hrs → 90 min design target for a first proposal draft

85 staff

The waiting room emptied before the doctor arrived

Four clinics moved patient intake to the phone in the waiting room and got eleven hours a week back at reception.

11 hrs/week target reception time returned across four clinics

Would Atlas fit your work?

Tell us what the business needs to do differently. If this platform is not the answer we will say which one is, or that the honest answer is none of them.

Scope an Atlas engagement