AI practice

The hard part of an agent is not the loop

It is deciding what it may do without asking. Persona, memory and skills are configuration. Tool permissions and the gates around irreversible actions are the design.

354

BACKEND TESTS, 13 SUITES

FRONTEND TESTS

ARCHITECTURE DECISION RECORDS

PUBLIC REPOSITORIES

The team

MVP to enterprise, hands-on

PremitiveKey runs an expert AI and machine learning engineering team with hands-on experience taking systems from a first MVP through to enterprise deployment. The work is practical rather than research: model and provider selection, retrieval design, agent orchestration, evaluation harnesses, and the deterministic code that has to sit around all of it.

MVP

Prove the idea works on real data before it is worth building properly. A narrow slice, a held-out evaluation set, and an honest read on whether the model earns its place at all.

Production

Tenancy, permissions, background processing, cost controls and observability. This is where most AI pilots stall, and almost none of it is model work.

Enterprise

SSO, audit trails, role-based access, data residency and approval gates on anything irreversible. One platform we built was engineered to ISO 27001 and taken through certification with the client.

Retrieval-augmented generationAgent and multi-agent orchestrationDocument extraction and KYCComputer vision and product matchingEvaluation and held-set scoringProvider abstractionPrompt and task contractsCost and latency budgeting

Four principles

How we build AI that can be trusted with real work

Each of these is visible in a public repository or a live product on this site, not just stated here.

The model reads and drafts. It does not decide.

Extraction, classification and drafting are language work. Scoring against a rule, computing a figure or taking an irreversible action belongs in deterministic code.

Withholding beats guessing.

A field the extractor is not confident about is returned blank and flagged, never filled with a plausible value. In retrieval, a relevance floor decides whether an answer is composed at all.

Humans approve anything irreversible.

A state in the workflow, not a convention in a prompt. The run pauses, persists, and resumes on a person’s decision, with the approval recorded.

Accuracy is measured, not asserted.

A held set of real cases, scored on every change, with precision and recall reported separately rather than as a single number.

Capabilities

What each one does, in plain language

Four capabilities with a system behind each. Three are public code; one is a recording of working software.

Capability	What it does	Behind it
Enterprise RAG	Answers questions from your own documents with citations, and declines when the documents do not contain an answer. Retrieval runs behind a relevance threshold, so the model is never called to guess.	EnterpriseRAG Built · public
Visual AI	A photograph replaces the search box. The system identifies each item, matches it to a catalogue by meaning rather than keyword, and returns nothing rather than a weak match.	Visual product discovery
KYC & onboarding	Identity documents extracted into typed fields. A field the model is not sure of comes back empty and flagged for a reviewer, never filled with something plausible.	KycCheckAI Built · public
Agent-driven QA	A model writes API test scenarios and keeps them current as the API changes. The suite runs on a schedule over a job store that survives a restart, so a regression is visible rather than discovered.	QAAPIAutomation Built · public

How we build agents · shown running

Two systems, traced step by step

The clearest answer to “how do you build agents” is to watch one work. EnterpriseRAG, our public retrieval platform, turns a document into searchable knowledge and a question into a grounded, cited answer. The agentic workflow takes a request through planning, evidence, review, human approval and execution. Each map is the real structure; the trace beside it records every step as it happens.

EnterpriseRAG · public repository

Enterprise RAG platform: from document to grounded answer

Documents are extracted, classified, chunked and embedded inside the organisation’s own infrastructure. Questions are planned, retrieved against the vector index, checked for grounding and answered with citations. 354 backend tests; source public.

Engineering study Source on GitHub

EnterpriseRAG · document to indexrun 1/2Built · public

Usersweb app · admin console · public chatbot

Employees & teamsweb application

Administratorsadmin console

Public userspublic chatbot

Identity providerSSO · OIDC, SAML

FrontendNext.js · TypeScript · route handlers as the BFF

Chat interfacegrounded answers, citations

Document managementupload, status, tags

Admin dashboardorganisations, settings, keys

Middlewaresecurity headers → request ID → CORS → auth → rate limit

JWT authenticationhttpOnly refresh, denylist

Rate limitingper IP, per user

Input & file validationtype, size, signature

Audit trailappend-only

Backend APIFastAPI · Python · SQLAlchemy

Document servicemetadata, duplicates, status

RAG & chat servicecontext, grounding, citations

Agent frameworkquery planner · response composer

LLM provider managementencrypted keys, fallback, routing

Workers & cacheCelery workers · Redis

Extraction & OCRPDF, DOCX, PPTX · Tesseract

AI classificationcategory, title, tags

Chunking & embedding1000 chars · local 768-d

Vector indexingdeterministic point IDs

Redistask queue, cache, limits

Data & modelsPostgreSQL · Qdrant · file storage · providers

PostgreSQLmetadata, chat history, audit

Qdrantdocument embeddings

File storageoriginals, OCR output

LLM providersOpenAI · Anthropic · Gemini · Groq

01RECEIVEdocument uploaded · web app or API
02VALIDATEtype, size and signature checked · duplicates rejected
03QUEUErecord written · processing task queued
04EXTRACTtext extracted · OCR only where a page has none
05CLASSIFYcategory, title and tags · via the LLM provider
06EMBED1000-character chunks · local embeddings, no API cost
07INDEXvectors upserted · deterministic IDs, safe to retry
08COMPLETEstatus COMPLETED · audit row written

Knowledge stays in-house. Extraction, chunking and embedding run on the organisation’s own infrastructure, and every vector has a deterministic ID, so a retry can never duplicate a document.

01ASKquestion submitted · last six turns loaded from cache
02ADMITtoken verified · per-user rate limit applied
03PLANquery planner refines the question · 25 s, no retry
04RETRIEVEtop-5 chunks above 0.35 relevance · completed docs only
05VERIFYevidence present · otherwise an honest refusal, no model call
06COMPOSEanswer written from the passages only · 60 s, one retry
07CHECKcitations verified against the retrieved chunks · PII masked
08ANSWERanswer with source references · message and sources stored

Grounded by design. The composing model sees only retrieved passages, every citation is checked against them, and when the evidence is not there the system says so.

Agentic systems · LangGraph, CrewAI, MCP

Agentic AI workflow: from request to audited action

The workflow, not the model, is in control: LangGraph sequencing with durable state, CrewAI agents under typed task contracts with explicit tool grants over MCP, a human approval gate on anything irreversible, and a trace of every run.

Architecture reference How we build agents

Agentic workflow · request to audited actionrun 1/2Reference architecture

Requestweb app · API · schedule · a person for approvals

Requestuser, API call or schedule

Human reviewerapproval console

Responsetyped result, returned

OrchestrationLangGraph state machine · CrewAI Flows

Workflow controlsequencing and routing

Durable statesurvives restarts

Retry & validation policyexplicit, per step

Human approval gatea state, on irreversible actions

AgentsCrewAI crew · typed task contracts

Plannerscopes the task

Research agentread-only tool grant

Analyst agentstructured output

Reviewer agentchecks the contract

Execution agentwrites only what is granted

Tools · MCPone declared contract per tool

MCP serverdeclared tool contract

Search & retrievalgrounded, cited

APIs & servicesinternal, third-party

CRM · ERPsystems of record

Modelsbehind an adapter, never wired through the app

LLMreasoning, drafting

Embeddingslocal or hosted

Provider adaptervendor is configuration

Data & observabilityPostgreSQL · tracing · audit

Application databasesystem of record

Logs & tracinginputs, calls, latency, cost

Audit trailappend-only, attributable

01RECEIVErequest enters the workflow · run ID, state created
02PLANplanner scopes the task · typed task contract issued
03RESEARCHresearch agent gathers evidence · read-only grant
04ANALYSEanalyst produces structured output · schema enforced
05REVIEWreviewer checks the output against its contract · pass
06APPROVEirreversible action waits for a person · approved
07EXECUTEexecution agent acts through the declared MCP contract
08RECORDrun traced: inputs, calls, latency, cost · audit row appended
09RETURNtyped result returned · state marked complete

The workflow is in control. Deterministic code decides what runs next, each agent does one job under a typed contract, and nothing irreversible happens without a person.

01WAKEthe schedule fires · run ID, state created
02PLANplanner scopes the task · typed task contract issued
03RETRIEVEresearch agent gathers evidence · read-only grant
04ANALYSEanalyst produces structured output · schema enforced
05REVIEWreviewer checks the output against its contract · pass
06UPDATEa reversible write through the declared MCP contract · inside its tool grant
07RECORDrun traced: inputs, calls, latency, cost · audit row appended
08RETURNtyped result returned · state marked complete

Unattended where that is safe. A reversible write proceeds on the reviewer’s pass; anything irreversible stops at the approval gate first. Either way the whole run is traced and attributable.

Choosing the model

Generative or predictive: the first decision, and the one most often wrong

Both are called AI. They are built differently, cost differently by two orders of magnitude, and fail differently. Getting this right at the start saves more money than any optimisation afterwards.

	Generative AI	Predictive AI
Produces	New content conditioned on context: an answer, a summary, a typed record, a draft	An estimate drawn from history: a class, a score, a rank, a number, an anomaly flag
Needs	A model, good context, and evidence to ground it. Little or no labelled data	Labelled history that resembles the traffic you will actually see
Right when	The input is unstructured and the output is language or a structured record. The rules are too many, or change too often, to write down	The question is closed-ended and you need the same answer every time, in milliseconds, at negligible cost
Fails by	Inventing something fluent and confident that is not true	Scoring well overall while being wrong on exactly the cases that matter
Measured by	Grounding, citation coverage, refusal rate, and a held set scored on every change	Precision and recall reported separately, on a held-out set, with drift watched after launch
Cost shape	Per call, ongoing, and it rises with usage: a design constraint from day one	Expensive to train once, then close to free per prediction
Our guard	A relevance threshold that stops the model being called when the evidence is not there, and a typed contract it must return	A confidence floor below which the system defers to a person rather than deciding

The pattern that works. Predictive first, generative last. The cheap, consistent model narrows, routes, ranks and flags; the expensive, variable one explains, composes and drafts, and only on what survived. Most of the AI rescue work we are asked to do turns out to be a generative model placed where a predictive one belonged, or a generative model with nothing in front of it deciding when to stay quiet.

Status

Where our AI actually runs

A current status of the AI systems we have built. We publish it because the discipline that keeps an agent out of production until it is trusted is the discipline you are hiring us for.

System	Status	Detail
MeshTribe · agent layer	Acceptance testing	Profile integrity, content and report triage, community editorial and relevance. Structured verdicts with reasons and confidence, an audit ledger, and human escalation on anything irreversible.
Cinepolis Indonesia · agent layer	Acceptance testing	Recommendation, demand prediction and re-engagement above a live booking platform.
Play Cinemas · agent layer	Acceptance testing	Two model providers behind one orchestrator, so the vendor is a configuration decision.
PolicyDesk · document AI	In-house	Insurer-aware extraction, renewal chains and family cover mapping for India’s insurance distributors.
EnterpriseRAG · retrieval	Built · public	Qdrant retrieval with 354 backend tests across 13 suites. Read the study.
Platforms beneath them	Live	MeshTribe, both cinema applications, FOXBOX and Sadhan are all in production.

No agent layer is live in production yet. The platforms are. We draw that line explicitly, on every page where a system is named.

Practice

How we work, in the open

8 repositories are public at github.com/PremitiveAI, and so are the architecture decisions behind them: 354 backend tests across 13 suites, 58 frontend tests, 8 architecture decision records and a 10-table schema you can read.

Architecture decision records

Eight published. Every significant decision is written down with its reason, so the next engineer inherits the thinking and not just the code.

Held-set evaluation

Accuracy is measured on a set of real cases, scored on every change, with precision and recall reported separately.

Typed task contracts

Prompts are provider-specific and swapping one changes the output. What works is defining your own tasks, each with a typed contract and its own test cases.

Read our code Open an architecture decision record

Go deeper

The architecture, the studies and the recordings

For a technical reader: the layered reference architecture, three engineering studies with public code behind each, and recordings of systems running.

Ask us where AI actually belongs in your product

Often the answer is a smaller slice than expected, and occasionally it is nowhere. You get that answer either way.

A 60-minute architecture review · no charge · the notes are yours either way