Case Studies
We help companies bring AI with ROI
Every number below is measured against a real baseline and scored on the client's own historical data. Where a system fell short, we say so — including what we changed as a result.
Showing 23 of 23 case studies.
Manufacturing — made-to-order fine jewelry
AI-powered vision-enabled agentic solution for ERP order intake automation
Custom orders arrived as free-text email with sketches and photos attached. Staff spent 15–30 minutes keying each complex order into Business Central. We built an intake agent that reads the thread and its images, fills ~25 job-card fields with calibrated confidence, and stops at a human approval gate.
Banking — 1M+ transactions daily
Fraud detection on a transaction knowledge graph
Rule-based screening caught 71% of fraudulent transactions and generated 40% false positives, with a 4–6 hour response time. We rebuilt detection on a transaction knowledge graph — 24,000 nodes, 176,000 edges — and layered graph neural networks and an LLM analysis interface over it.
Blockchain — smart contract developer tooling
Knowledge-graph tuning for a smart-contract copilot
A smart-contract development copilot needed to generate compilable, test-passing Solidity. The gains came from three places that are not the model: migrating the live knowledge graph to Memgraph, hand-refining the Solidity ontology with human-in-the-loop review, and adding a post-processing pass that makes raw output SolEval-compliant.
Marketing — social content production
Multi-agent content production copilot
The client was spending $480K a year and roughly six hours per carousel on research, writing and revision, and still missing trending topics to slow research cycles. We built a four-agent workforce — retriever, writer, reviewer, coordinator — that produces a finished carousel in eight minutes.
Enterprise technology — support and field engineering
Knowledge-graph RAG for technical documentation
Support and field engineering teams at an enterprise technology company could not find information across thousands of technical documents, and keyword search hid the connections between related deployment scenarios. We ingest any document format, build a knowledge graph that preserves relationships across document boundaries, and answer in natural language.
Telemarketing — Vietnam, multi-dialect
GraphRAG voice automation for multilingual telemarketing
A Vietnamese telemarketing organisation was drowning in repetitive high-volume calls, with inconsistent service across regions and product knowledge fragmented across PDFs, wikis and spreadsheets. We built a model-agnostic voice platform grounded in GraphRAG, with accent-stratified analytics and confidence-based human handoff.
Payments technology — APEC region
LLM campaign constructor for regulated marketing
Marketing teams were bottlenecked on manual translation and editing, and every piece of copy carried regulatory exposure. We built an LLM Constructor inside their existing campaign portal: templates and parameters for language, tone and channel, with regulatory checks embedded in the prompts and post-generation review logged for audit.
Financial services — security operations
RAG document classification for security operations
The client’s security analysts were doing document admin instead of security — manual classification in Confluence, hours lost searching, inconsistent labels across thousands of files. We built a RAG system that classifies zero-shot from category descriptions alone, plus semantic search over the corpus.
Agriculture — large-scale livestock production
Knowledge-graph precision agriculture
The client fed entire barns based on what most animals need, using fragmented data across hundreds of farms and manual spreadsheet decisions. We mapped every relationship — animal genetics, feed type, environment, health score, logistics — into a knowledge graph, and made GraphRAG the interface to it.
Travel and tourism — national tourism board
Multi-agent itinerary orchestration
A national tourism board needed to coordinate 50+ vendors in real time with the judgment of a local expert — not static recommendations. We built an AI conductor orchestrating specialised agents over a shared scratchpad, so a change to one booking re-plans the rest against its real constraints.
Food manufacturing — multi-plant pork processing
Order placement optimisation solver
Manual order placement across five plants cannot weigh yield variance against labour rate against freight against available capacity, per order, at volume. The solver does, under the real production and inventory constraints — and replays completed orders to show what each one would have earned at the plant it should have gone to.
Financial services — equity capital markets
Multi-agent automation for equity capital markets
Deal flow, issuer data and market commentary lived across spreadsheets, PDFs and a deal database, and every memo, weekly overview and IOI update was assembled by hand. We built a chat-style ECM interface where a supervisor agent routes each command to a specialist — deal analysis, document generation, market intelligence, alerts — all reading one knowledge graph of the deal history.
Financial services — consumer lending contact centre
Automated QA for every contact-centre call
Quality assurance in a large bilingual contact centre meant reviewers listening to a hand-picked 2–3% of calls and scoring them by hand. We built an automated quality management system: speech recognition at 80%+ accuracy, agents that score emotion, script compliance and conversation flow against the QA team’s own criteria, and a supervisor agent that turns every score into coaching — feeding the CRM, data warehouse and workforce management tools.
Medical devices — aesthetics and physiotherapy equipment
Lead qualification and device support agents
Clinics and practitioners wrote in over WhatsApp and the website with everything from demo requests to device error codes, and each message waited for the right person. We built a multi-agent ecosystem on Agno: a triage agent routes every message to a sales assistant that qualifies leads and books demos, a customer FAQ agent that answers in the customer’s language, or a technical support agent that troubleshoots from the manuals and logs tickets.
Healthcare & life sciences — dermatology field sales
Voice-first assistant for pharmacy field sales
A dermatology company’s 600+ field sales reps prepared each pharmacy visit by hand across a data warehouse, a CRM and an ERP, then documented it afterwards. We built a voice-first assistant with two agents: one that speaks a pre-visit briefing in under five seconds, and one that turns a spoken visit account into CRM updates and follow-ups in under three — every query and write filtered to the rep’s own territory.
Education — online assessment preparation
Real-time AI avatar benchmark for interview practice
A live interview avatar has to respond in under a second, look professional and cost little enough to run for every practice session. Vendor demos answer none of that on your own infrastructure. We deployed three pre-selected avatar candidates on SageMaker, ran each on T4 and V100 GPU configurations, measured every setup the same way and delivered one recommended solution-plus-hardware pairing with a production roadmap.
Financial services — consumer fintech
Agentic text-to-SQL for business users
Business teams in sales, logistics and operations waited on analysts for every number. We built a ReAct agent on Agno that reads the schema and house terminology, generates SQL, passes it through a rule-based validator, runs it over a private link with row-level security and returns a plain-language answer in chat — with MCP connecting it to the data and encryption keys held by the client.
Healthcare & life sciences — consumer health and nutrition
Personalised meal planning on a nutrition knowledge graph
User profiles, health-test results, recipes and ingredients lived in separate places, and generic meal suggestions ignored both the test results and the restrictions. We connected them in one nutrition knowledge graph and put Graph-RAG agents on top, so every meal plan is built from the user’s own subgraph and explains why each meal is there.
Healthcare & life sciences — patient access to pre-approval treatments
Regulatory monitoring agents for pre-approval treatment access
Expanded access, compassionate use, orphan drug and clinical trial rules change across registries, government sites and publications, and keeping track meant people reading them. We built a multi-agent system that ingests those sources, detects and classifies changes, validates them with an LLM judge and a human checkpoint, and answers pathway questions from a knowledge graph with the source cited.
Cybersecurity — governance, risk and compliance consulting
AI attack path analysis for a GRC consultancy
Mapping how an attacker could move through a client’s infrastructure, scoring the impact and tying each finding to ISO 27001, NIST or NIS2 controls was done by hand, engagement after engagement. We built a bilingual copilot where a triage agent routes each question to risk analysis, impact assessment, compliance mapping and recommendation agents, all reasoning over one semantic knowledge graph.
Financial services — insurance
Photo and video verification for insurance
Customers submitted vehicle photos, 360-degree videos and VIN close-ups for verification, and a team reviewed each one by hand against policy rules. We built a vision-language verification module that rates image quality, confirms each shot shows what was asked for, decodes the VIN and flags discrepancies with the policy record — with instant feedback to the customer when a reshoot is needed.
Financial services — wealth management
AI assistant for wealth relationship managers
Market research, house views and fund data were spread across PDFs, chat threads and inboxes, and every client deck and meeting note was assembled by hand. We built one RM assistant over a knowledge map of the bank’s own research: cited answers in seconds, talking points from overnight news, three-slide fund comparisons, a daily list of who to call, and meetings that log themselves.
Financial services — commercial banking
AI customer support for a commercial bank
Routine questions about policies, procedures and accounts sat in the same queue as the matters that needed a banker. We built a multi-agent support system where a triage agent routes each inquiry to a specialist — general banking, account support, technical support or escalation — all answering from a GraphRAG knowledge graph of the bank’s normative documents, procedures and FAQs, and handing complex cases to a person with the full conversation attached.
Have a process that runs on email and hand-keying?
The pattern generalises: an unstructured request arrives from someone who knows their trade but not your field names, and a person turns it into a structured record by hand.