PremitiveKey
AI systems & product engineering

AI development,
engineered for
production

AI agent development and agentic AI systems, SaaS platforms, web and mobile app development, MVPs and enterprise software — designed, built and taken to production by one engineering team.

Book an architecture review See the architecture
EnterpriseRAG · query trace run 1/3 Built · public
  • 01RECEIVEquestion · from the chat UI
  • 02PLANquery planner · 25s budget · no retries
  • 03RETRIEVEvector search · top-K 5 · filter in query
  • 04VERIFYevidence checked · relevance 0.35 gate passed
  • 05COMPOSEanswer returned · with citations
  • built-in accuracy safeguard
  • 06PROTECTunverified answers blocked · nothing is ever guessed
Grounded by design. Every answer is composed from verified sources and cited — accuracy is engineered in, not hoped for.
  • 01WAKEscheduled run · unattended
  • 02GENERATEtest scenarios · as the API changes
  • 03RUNsuite executes · over a durable job store
  • 04SURVIVErestart mid-run · the job store persists
  • 05REPORTregressions visible · not discovered
Unattended is the point. The suite runs on a schedule, the job store survives a restart, and regressions become visible rather than discovered.
  • 01RECEIVEidentity document · image or PDF
  • 02EXTRACTtyped extraction · schema-bound
  • 03VERIFYevery field checked · confidence scored per field
  • 04ROUTEuncertain fields to an expert · nothing guessed
  • 05CONFIRMreviewer approves · every value trustworthy
Accuracy first. Uncertain fields go to an expert instead of being guessed — so every stored value can be trusted.
50+
Products delivered
100+
Client portals in production
30
Brands on a platform we built
8
Public repositories
Every system on this site carries its actual state Live Acceptance testing In build Built · public — and no agent layer is called live until it is.
What we build

AI development, and the software it lives inside

Agentic AI development, AI agent development and applied AI engineering, plus the SaaS, web, mobile and custom software development around it. One team across all of it.

AI agent developmentA defined role with tool access and explicit permissions — reached through MCP or a typed tool contract — built for bounded, repeatable judgement work inside a real workflow.
Agentic AI developmentMulti-step agentic workflows and multi-agent systems on CrewAI, LangGraph and LangChain: conditional routing, integrations, validation and orchestration held in deterministic code.
Generative AILLM systems that produce something — an answer, a summary, a typed record, a draft — from unstructured input, with grounding, citation and a refusal path.
Predictive AIClassification, scoring, ranking, forecasting and anomaly detection: closed-ended questions answered consistently from history, with precision and recall reported separately.
Retrieval and RAGChunking, embedding, hybrid vector and keyword search, reranking and metadata filtering — with a relevance threshold that decides when the LLM should not be called at all.
Data and document pipelinesIngestion, OCR and vision extraction, validation, typed persistence and indexing — the layer that decides whether anything above it works.
MVP developmentA first working version built to test the idea against real users and real data, on a foundation that does not have to be thrown away.
SaaS developmentMulti-tenant platforms with identity, permissions, billing, audit and the AI capability designed in rather than added later.
Mobile app developmentNative and cross-platform applications shipped to the App Store and Google Play, over a shared backend contract.
Web and software developmentWeb applications, APIs, backend services and system integrations — including work with no AI in it at all, where that is the right answer.
Enterprise software developmentEnd-to-end systems connected to CRM, ERP, document stores and internal services, under existing access control and audit requirements.
Cloud, DevOps and CI/CDAWS, Azure, Google Cloud and DigitalOcean. Containers, pipelines that build and deploy on every merge, infrastructure as code, monitoring and cost control.
All fourteen capabilities Cloud & DevOps
Generative and predictive AI

Two different tools. Most projects need both.

“AI” covers two families of model that behave nothing alike, cost nothing alike, and fail nothing alike. Choosing between them is the first real decision on any project, and it is the one most often got wrong.

Generative AI

Produces something that did not exist

An answer, a summary, a typed record pulled out of a scanned document, a draft, a classification with its reasoning attached. Right when the input is unstructured, the output is language or a structured record, and the rules are too many or too fluid to write down.

Retrieval & RAGDocument extraction SummarisationDrafting Conversational interfacesCode generation

Where it is expensive to get wrong: a generative model will produce a fluent, confident, entirely invented answer unless you engineer against it — grounding, citation, typed output contracts, and a relevance threshold that stops the model being called when the evidence is not there.

Predictive AI

Estimates an outcome from history

A class, a score, a rank, a number, an anomaly flag. Right when you already hold labelled history, the question is closed-ended, and you need the same answer every time — cheaply, in milliseconds, and measurably.

ClassificationForecasting Ranking & recommendationAnomaly detection Risk & propensity scoringSegmentation

Where it is expensive to get wrong: a model that scores well in aggregate can still be useless on the cases that matter — so precision and recall are reported separately, drift is watched after launch, and below a confidence floor the system defers to a person.

How they belong together. The predictive model is cheap, fast and consistent, so it goes first — narrowing, routing, ranking, flagging. The generative model is expensive and variable, so it goes last — explaining, composing, drafting, and only on what survived. Reaching for a generative model where a predictive one would do is the most common and most costly mistake we are asked to correct.
Selected work

Software we designed, built and support

Products delivered end to end, from architecture through to production — each carrying its actual state. The case studies carry the challenge, what was built, and the numbers.

FOXBOX Live

A multi-tenant rewards, gifting and engagement platform: four HUB products on one codebase, 100+ live client portals. Built for FOXBOX, with PremitiveKey as the end-to-end technical partner — FOXBOX owns and operates it. Engineered to ISO 27001 and taken through certification with the client.

100+
Client portals live
10+
Countries served
30
Brand programmes
1M+
Shipments a year

Sadhan

India’s first 100% regional multi-lingual stock market application — a trading platform and an investor knowledge centre in one product. Delivered across web, mobile, cloud and DevOps. Live on Google Play.

Capital marketsLive

MeshTribe

The digital home for motorcycle riders: 8 product surfaces on a single backend of 22 domain modules, live on web, iOS and Android. An AI agent layer above it is in acceptance testing.

CommunityLiveAgents · acceptance testing

Spacemark

A centralised banking system for cooperative banks — daily deposit collection through 130 field agents, reconciliation and one system of record. Over ₹80+ crore processed.

BankingLive

Jaldicash

A pan-India assisted-payments platform for Weizmann Group: domestic money transfer, AEPS, recharge, bill payment and travel ticketing on one login, with ledger reporting.

PaymentsLive

Cinepolis Indonesia & Play Cinemas

Booking applications for two cinema chains — showtimes, live seat inventory, payment and ticketing across iOS and Android. Delivered with KRS Infoserve. Agent layers in acceptance testing.

TicketingLiveAgents · acceptance testing

PolicyDesk

A daily action list for India’s insurance distributors: policy documents are read automatically, renewals are linked into one chain per client, and gaps in a family’s cover are surfaced before they become missed business.

InsuranceIn-house

LocalKhoj

A hyperlocal discovery platform connecting a community to the businesses around it — verified listings, categories and search by locality, ratings and reviews, and offers published by the businesses themselves.

Local discoveryIn-house

Captis & Somaiya Vidyavihar

An enterprise e-learning platform delivered for CyberNX Technologies, and the university website and content management platform for Somaiya Vidyavihar.

Education

Maharashtra Police Housing Corporation

A KRA performance system for construction engineers across the whole of Maharashtra — goals, ratings and accountability in one system of record.

Government · KRA system
All case studies
FOXBOX · platform reach

Brands running on a platform we engineered

FOXBOX is owned and operated by FOXBOX. It is their product and their business. PremitiveKey was the technical partner that designed, built and continues to support the platform. The brands below are FOXBOX’s customers, not PremitiveKey’s — shown because they evidence the load, the scale and the compliance expectations the platform we engineered has to carry.

Amazon
Deloitte
Uber
Morgan Stanley
verizon
General Mills
Merck
Hitachi Vantara
WRTH
NIQ
GlobalLogic
SKF
Automation Anywhere
NetApp
OakNorth
Gainwell
ICICI Bank
Tata Mutual Fund
Kotak Life
LT
Persistent
Hexaware
CaratLane
Pine Labs
Darwinbox
Indegene
Delhivery
InCred
Amity
Sika India
Architecture · CrewAI, LangGraph, MCP

How the systems are actually put together

A production AI system is mostly not the model. It is the workflow that controls execution, the tools that reach real business systems, and the data and observability layer that makes the result auditable.

Read the engineering study Full architecture reference
Reference systems · shown running

How these systems run, step by step

Two reference architectures traced live. EnterpriseRAG, our public retrieval platform, turns a document into searchable knowledge and a question into a grounded answer. The agentic workflow takes a request through planning, evidence, review, approval and execution. The trace beside each map 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.
Demonstrations · agentic systems

Recorded from working software

Each recording shows agents doing bounded work inside a real workflow. Every agent has a defined role, tool access it was explicitly granted, a typed output contract, and a fallback for the step where it cannot decide.

Vision agent · retail

Visual product discovery

An image replaces the search box. A vision agent identifies each worn item, a retrieval step matches it to the catalogue by meaning rather than keyword, and the result carries price and nearby availability. Where the match is weak it returns nothing rather than a guess.

Multi-agent workflow · fashion

Fit-to-model try-on

Real-time virtual try-on against a body profile, then a recommendation step for style, size and fit. Generation and recommendation are separate agents with separate contracts, orchestrated in deterministic code. Built for an enterprise client in fashion retail.

Scheduled agent · engineering

API QA automation agent

Generates and maintains API test scenarios as the API changes, on a schedule and unattended, over a job store that survives a restart. Deployed fully on-premise. Phase one complete, phase two in progress.

All demonstrations
How an engagement starts

Three steps, and the first two cost nothing

No discovery fee and no workshop invoice. The way to find out whether we are the right team is to put a real problem in front of us and read what comes back.

STEP 01 · FREEA 60-minute architecture review

Bring the system you have or the one you are planning. You get our honest read on where the risk sits and what we would decide differently.

STEP 02 · YOURS TO KEEPThe notes, in writing

What we would build, what we would not, and why — useful whether or not we ever work together.

STEP 03 · ONLY IF IT MAKES SENSEA scoped engagement

If the review points to real work, we scope it to the problem in front of us. If it does not, we say so.

Put a real system in front of us

Sixty minutes on your architecture, your constraints and where AI does and does not belong in it. No charge, and the notes are yours either way.

Contact us Read our code
A 60-minute architecture review · no charge · the notes are yours either way