Agentic AI workbench
An AI assistant that safely uses real tools (files, terminals and multi-step workflows) with permissions and a full activity log. 3,000+ automated tests gate every release.
Home Work
ProofSystems our team has actually built.
AI agents, document and voice AI, data automation, products and cloud work, from production systems to research in progress. Filter by what you need.
Appraisal review engine
A RAG system that reviews commercial appraisal reports against USPAP and FIRREA criteria. It pairs LLM review with deterministic maths and sensitivity checks, and writes PDF and JSON reports.
Urdu speech-to-text
Whisper fine-tuned for Urdu, served through REST APIs that feed real-time dashboards and call prioritisation on a public-sector platform.
Production OCR model
An OCR model in operational use on a lending automation platform for banks and credit unions.
Handwritten prescription reader
Vision-language AI that extracts text from handwritten eye prescriptions.
CPU-only document extraction
Layout detection, quality filtering and lightweight OCR that process structured documents on standard hardware, with no GPU.
Visual search engine
Identifies a movie from a single frame using CLIP embeddings, vector search and AI re-ranking.
Vehicle counting
Categorical vehicle-counting pipelines built on YOLOv8.
Public-records automation
Three production systems for real estate clients. They extract, validate and enrich public records, then deliver scheduled reports with no one watching.
Unattended data pipeline
Scrapes permit records, enriches them with property data and emails weekly reports, with retries, gap-filling, circuit breakers and health alerts.
Call-volume forecasting
Time-series models that anticipate call volume and incident trends, for resource planning and prioritisation.
Outcome prediction from early results
A model that predicts final results from early assessments, packaged as a reproducible MLOps pipeline.
Contract risk analyser
Checks 17 types of risky clauses, explains every flag and keeps working offline.
Domain-tuned language models
A 9,600-example dataset; Llama 3.1 8B and TinyLlama fine-tuned and published with held-out evaluation.
Generative-media SaaS
An AI image and video generation platform with accounts, Stripe subscriptions, usage credits and cloud deployment.
AI learning platform
Quizzes generated from uploaded material, an AI study assistant and XP progression.
Dispatch & client portals
Real-time dispatch with SMS notifications and client-facing status portals.
Rental operations platform
Orders and inventory tracked across several warehouses, product lifecycle statuses, Stripe payments and QuickBooks invoicing.
Multi-store inventory sync
Parent and child Shopify stores kept in sync, with subscription billing.
Production monitoring
Production systems watched with Prometheus, Grafana, Kibana and Nagios, with root-cause analysis and SLA reporting.
LLM-based Kubernetes autoscaling
Research on LLM-driven autoscaling and performance-aware scheduling for microservices, with OpenTelemetry observability.
Number-plate recognition
CPU-first recognition for Latin and Arabic plates, built on 723,887 audited images.
AI evaluation at scale
A 347,236-cell study across four AI model families on 12 parallel GPU workers.
FBR and PRA connectors
Invoice connectors built from FBR's DI API v1.12 and PRA's fiscal-device specification, tested against every documented response.
Restaurant billing and ordering demo
A working demo: bills with a PRA-style receipt and QR code, a WhatsApp order flowing into the POS, and an owner dashboard.
Try itSystems built by members of our team, including work delivered in other roles; employers and client names are withheld. "In our lab" marks research and products in progress. "Concept demo" marks working samples, not client projects.
Your project
Have something like this in mind?
Tell us what you want to build. The first call is free and takes 30 minutes.
hello@operandalabs.com
Lahore · Dubai · Finland · working worldwide