AI systems that survive
contact with production.

Agentic AI, RAG and automation — built to be checked, not trusted.

See how we built it Start a conversation
Scroll

Most AI pilots die between the demo and deployment.

The agent works beautifully in a notebook, then hits real data and starts hallucinating. The RAG pipeline degrades. Nobody can explain what the model decided. Security has no audit trail.

The project quietly gets switched off.

We build for the other side of that gap.

Featured work — Scopeline, AI-assisted estimating & document review

An AI tool found low-six-figures of unpriced scope four days before a tender closed — on a package the team had already read by hand.

The client

A mid-size mechanical and plumbing contractor bidding institutional, commercial and industrial work in Southern Ontario. Their five-person estimating team was losing roughly a third of every bid cycle to document handling rather than pricing. Named client withheld under NDA.

The problem

A typical tender package arrives as 900–1,600 pages of mixed PDF, plus a drawing set of 180–380 large-format sheets. Equipment schedules live as vector tables inside those sheets — not as text, not as a spreadsheet, and frequently rotated 90°. Specification-to-drawing verification was done by memory and highlighter, under deadline, by whoever was free.

What we built

Scopeline ingests an entire tender package and produces three reviewable artifacts in about forty minutes: a contract risk register, a structured equipment and valve schedule extract, and a scope discrepancy report comparing what the specifications require against what the drawings actually show.

Every extracted number, clause and schedule row carries a provenance link back to a page, a bounding box and a source span. An estimator can click any figure and land on the exact region of the exact sheet it came from. That single design decision is what moved the tool from an interesting demo into something the team bid live work with.

The engineering note

Never ask a vision model to read a whole drawing. An E-size sheet downsampled enough to fit a single call renders schedule text at two or three pixels of stroke height. The model will still return numbers — they just won't be the right ones.

Region-first reading of a large-format drawing sheet Left: a whole E-size sheet downsampled into one model call, where schedule text collapses to two or three pixels. Right: deterministic analysis proposes candidate regions, and only those regions are read at native resolution. one call · whole sheet schedule text ≈ 2–3px stroke · numbers returned, not read regions proposed read at native resolution accuracy on schedule rows ≈ 2×
Cheap deterministic analysis decides where to look; the model only reads what it can actually see.

Instead: cheap deterministic analysis proposes regions — vector ruling lattices, text-density clustering, whitespace gutters — and the model reads those regions at native resolution. Accuracy on schedule rows roughly doubled after that change.

Then three checks before anything is accepted:

  • Row count must match detected horizontal rulings.
  • Values must fall in trade-standard ranges.
  • The same tag on two sheets must resolve to the same record.

An illegible cell is reported as illegible. It is never inferred from the rows above it.

40m

Package receipt to reviewed contract risk register — down from 6.5 hours

0.94

F1 on structured schedule extraction, 1,100-row gold set

97%

Recall on the 42-item high-risk clause taxonomy

0.6%

Silent error rate at handover — down from 12.4%

Measured on the client's own tender packages, from first integration through handover.

Also delivered — cross-platform mobile product

A privacy-first dating app, built for Android and iOS

Meeting people without handing over your identity

A location-based social product for Android and iOS where people meet without exposing their name, number or exact location. Built from one Flutter codebase, with a FastAPI backend, a separate real-time chat service, and a web dashboard for moderation and analytics. Delivered tested and ready for store submission.

Location is handled as map zones rather than GPS coordinates — exact position is never stored, and zones below a minimum population are hidden entirely. Photo uploads have their embedded location data stripped on the device before anything is transmitted.

Built with Flutter · Python · FastAPI · PostgreSQL + PostGIS · Celery · Firebase · Cloudflare R2
  • 84 features delivered
  • 1,240 automated tests passing
  • 99.8% crash-free sessions
  • 1.9s cold start on a mid-range handset
  • 20 physical devices tested

Anatomy

Every system we build has the same spine.

The interesting part is not the model call. It is the validation step before an action fires, and the log that lets you reconstruct why it fired.

Automations

  1. Trigger event
  2. Data ingestion
  3. AI processing
  4. Decision and validation
  5. Action and integration
  6. Logging and monitoring
  7. Notification

Triggers

EmailWebhook or APIForm submissionFile upload

AI processing

ClassificationEmbeddings and searchExtractionSummarisation

Actions

Send replyGenerate reportCreate taskUpdate CRM

Monitoring

LogsAnalyticsAlerts

A trigger enters at the top, the model does one narrow job, an action fires only once it passes validation, and every step writes to the log. Swap the trigger and the action and the middle stays the same, which is why a second automation costs a fraction of the first.

Mobile apps

The same shape with one difference that changes the architecture: the handset has to keep working with no signal. Anything that can run on the device does, and the sync queue drains when the connection comes back.

  1. Interface
  2. State
  3. Online check
  4. Local store or sync queue
  5. Cloud and model
  6. Update the UI
  7. Analytics
  8. Push notification

What the person does

Camera captureManual entryLocationPush notifications

On the device

Local storageOffline queueMetadata strippingSecure session

In the cloud

Firebase or SupabaseAPI and model callsGeospatial queriesMedia storage

What comes back

Cloud syncInsightsChartsNotifications

The handset has to keep working on a poor connection, so state lives on the device first and syncs when it can. Anything sensitive — exact location, photo metadata — is stripped before it leaves the phone.

Web platforms

A request comes in, authentication decides what it is allowed to touch, and the model is one service among several rather than the centre of the system. Most requests never reach it.

  1. Interface
  2. API layer
  3. Auth check
  4. Business logic
  5. Model or database
  6. Cache
  7. Response
  8. Update the UI
  9. Logging

Frontend

Next.js and ReactState managementTailwind CSSReal-time WebSockets

Backend

Node.js or FastAPIREST and GraphQLJWT and OAuthBackground jobs

AI layer

LangChain orchestrationModel callsVector searchSpeech and vision

Data layer

PostgreSQLMongoDBRedis cacheObject storage

Integrations and outputs

StripeDashboards and exports

Background jobs carry anything slow, so a model call never blocks a page render. The cache sits in front of the expensive paths, and every request writes to the log whether it touched the model or not.

Solution designs

Eight systems, designed before a line of code.

Each is a complete design pack — scope, architecture, guardrails, running costs and a gated build plan. It's the part most teams skip, and it's where projects are actually won or lost.

AI agents & automation

Agentic Workforce for a Marina & Boatyard

Seven narrow AI agents covering the front desk, phone line and the weeks-long chase for deposits, contracts and insurance certificates.

Claude Sonnet + HaikuTypeScript orchestratordurable workflowsMCP tool layerTwilio VoicePostgresLangfuse
Event inOrchestratorNarrow agentGuardrail layerApprove or act

Read the design

AI agents & automation

Unified Lead & Conversation CRM

Leads and messages from Facebook, Instagram, Google Ads, WhatsApp, web chat and email in one inbox — with an AI assistant that replies when no agent has within five minutes.

FastAPI or NodeReact + WebSocketPostgres + pgvectorSQSEventBridge SchedulerAmazon BedrockECS FargateTerraform
WebhookQueueResolve identity5-min timerAgent or AI

Read the design

AI agents & automation

AI Phone Agent

An inbound voice agent that holds a natural, interruptible conversation, qualifies callers, books appointments in the CRM and calendar, and hands off to a person when it should.

Twilio Voice + Media StreamsOpenAI Realtime APIbridge serverfunction callingCRM + calendar integrations
CallMedia streamRealtime modelActionsHuman handoff

Read the design

Mobile & marketplaces

Customer Property Portal

A mobile app that lets a real estate developer's buyers follow their property from booking to handover — construction progress, payments, documents and service requests.

FlutterSupabase / managed PostgresFirebase Cloud Messagingback-office web app
Site updateBack officeMilestoneNotificationCustomer app

Read the design

Mobile & marketplaces

On-Demand Hairstyling Marketplace — MVP

A two-sided marketplace where customers book a mobile stylist to come to them, stylists manage services and availability, and an admin panel approves who joins.

FlutterSupabase Auth + RealtimePostgres + PostGISFirebase Cloud Messagingadmin web panel
BookMatchStylist acceptsLive statusComplete

Read the design

Mobile & marketplaces

Restaurant Discount Marketplace

Customers find and claim nearby restaurant deals; restaurants run last-minute offers and scan redemptions at the counter; an admin panel keeps both sides honest.

FlutterNode + TypeScript or PythonPostgres + PostGISRedisFirebase Cloud MessagingReact admin
DiscoverClaimQR codeStaff scansRedeemed

Read the design

Platforms & data

Home Care Training & Compliance Platform

A mobile-friendly web platform that helps home care agencies track and prove annual staff training — courses, quizzes, e-signatures, trainer sign-off and audit-ready reports.

Next.jsmanaged PostgresTwilio SMS + emailcertificate generationresponsive web
Hire dateHour requirementCourse + quizTrainer certifiesAudit report

Read the design

Platforms & data

Market-Event Monitoring Service

A read-only service that tells trading software whether a ticker, or the whole market, is inside an event window — earnings, FOMC, CPI, halts — and when normal processing can resume.

Event collectorsPostgres append-only event storein-process state cacheSEC EDGAR + economic calendarsAWS
CollectorsNormaliseEvent logWindow stateQuery or stream

Read the design

What we build

Three things, done properly.

Applied AI

Agentic systems and multi-agent orchestration. RAG applications and enterprise knowledge assistants. Voice and calling agents. Chatbots and conversational automation.

LangChain · LangGraph · vector retrieval · model fine-tuning · OpenCV

Product engineering

The full product around the model — APIs, dashboards, mobile apps, and the unglamorous plumbing that makes a model usable by people who don't care that it's a model.

Python · FastAPI · Django · Flask · React · Next.js · Flutter · PostgreSQL

Deployment & AI security

Getting it live and keeping it defensible. Cloud infrastructure, CI/CD, monitoring, and controls at the model, data and agent layer — permissions, audit trails, and guardrails against prompt injection.

AWS · Azure · Kubernetes · Docker · SSO · single-tenant deployments

Stack

AI & machine learning
LangChain · LangGraph · Claude · OpenAI Realtime · vector retrieval · pgvector · OpenCV
Backend & data
Python · FastAPI · Django · Flask · Node.js · PostgreSQL · PostGIS · Redis · Celery · Supabase
Frontend & mobile
React · Next.js · TypeScript · Flutter · Firebase
Automation & integrations
Twilio · Meta Graph API · Google Ads API · Stripe · Gmail
Infrastructure & security
AWS · Azure · Kubernetes · Docker · Terraform · Vercel · CI/CD · SSO · single-tenant VPC deployments · audit logging

How we work

Three stages, in this order.

  1. Discovery

    We read your documents, your data and your constraints before proposing anything. Where accuracy matters, we build the evaluation set first.

  2. Working prototype

    A narrow end-to-end slice, running on your real data, early enough that integration problems surface while there's still time to fix them.

  3. Hardened deployment

    Evaluation gates, monitoring, cost controls, runbooks and documentation. You own the code and can operate it without us.

Got an AI pilot stuck in pilot?

That's the conversation we want to have. Tell us what you're trying to build and what's blocking it.

Email us LinkedIn
contact@aprellasolutions.in Noida, Uttar Pradesh, India