Customers

The work, and what it changed

Selected engagements across education, recruitment, finance, and industry. Each entry names the problem, the system we built, and the results as the client measured them.

employees group-wide: 500+; employees group-wide
clients: 60+; clients
countries delivered into: 10+; countries delivered into
delivering since: 2016; delivering since

Outcomes at a glance

The results, before the detail

One signature result from each engagement below, each in the unit that mattered to the client. The detail follows.

Financial services

99.95%

platform uptime, under a Swiss data-residency SLA

40+ hrs/week reconciled by hand15 min/day, automated

Swiss multi-bank platform

Enterprise sales

qualified pipeline, with a 40% shorter sales cycle

Ricoh sales agent

Recruitment

2h to 15m

per interview, evaluator consistency up to 95%

Interview evaluation

Architecture

95%

match accuracy against the manual baseline

Wood-pattern visual search

Education

A week to hours

grading cycle, 90% of student queries auto-resolved

Edu Insight

Technology services

4 months

to a full auto-deploying evaluation pipeline

RAG evaluation system

Each figure is a real outcome from the engagement named on its tile. Results are shown in their own unit, not ranked against one another.

Featured engagements

Measured at the workflow, not the demo

Financial services

Switzerland

Azure Data Factory
Azure Databricks
TimescaleDB
Swissquote AMI v2
QuantLib
React + TypeScript
b.link Open Banking

One platform across eight banks, with Swiss data residency

The problem: A Swiss wealth management firm ran portfolio operations across 8+ global banks by hand: 40+ hours a week of manual reconciliation, discrepancy detection measured in days, and no real-time view of client positions, all under strict Swiss data residency rules.
What we built: A multi-bank data platform on Azure Switzerland. Swissquote AMI v2 and b.link Open Banking feed a harmonization layer; a QuantLib calculation engine computes TWR/MWR, risk, and volatility into TimescaleDB; a React and TypeScript front end serves real-time monitoring behind immutable, geo-fenced audit trails.

99.95%

platform uptime SLA

150,000+

portfolio calculations a day

<100 ms

response time across portfolios

banks integrated on one platform

Manual reconciliation fell from 40+ hours a week to 15 minutes a day, discrepancy detection from 3-5 days to under 5 minutes, and data accuracy from 94.2% to 99.95%. 6 members, 6 months.

Enterprise sales

Australia

LangChain
OpenAI GPT
Pinecone
Salesforce SDK
AWS Bedrock
AWS Lambda
HuggingFace

A conversational sales agent for Ricoh Global

The problem: Ricoh, a global enterprise, needed to automate lead qualification, prospect engagement, and pipeline management, targeting a shorter sales cycle and more qualified pipeline through AI-driven conversations tied to CRM.
What we built: A conversational AI engine handling objection handling and autonomous meeting scheduling, a Salesforce SDK integration layer syncing lead data and pipeline stages in real time, a lead-scoring ML pipeline for persona-based multi-channel outreach, and a sales analytics dashboard for forecasting and conversion metrics.

40%

faster sales cycle

qualified pipeline

Real-time

lead scoring on win probability

Multi-channel

outreach across email, LinkedIn, and CRM

Ricoh, Global Enterprise (Australia). 5 engineers, 4 months.

Technology services

Confidential

Python 3
RAG Pipelines
AWS Bedrock
OpenSearch
AWS Neptune
Docker
ReactJS

Automated evaluation for RAG pipelines

The problem: An LLM-specialized client needed a systematic way to evaluate and optimize Retrieval-Augmented Generation pipelines across different retrieval strategies and data sources. Manual benchmarking was slow and inconsistent, and it was blocking production deployment.
What we built: A modular RAG architecture with pluggable retrieval and LLM components, an evaluation engine scoring faithfulness, answer relevance, and context precision per run, and an auto-deploy step that promotes the best-performing configuration straight to the client chatbot each cycle. Built on AWS Bedrock and OpenSearch.

Automated

end-to-end benchmarking

scoring metrics per run

Auto-deploy

best config to the chatbot

4 months

to full system delivery

3 members, 4 months, 5 features. Scored on faithfulness, answer relevance, and context precision.

Architecture and interior design

Confidential

Vision Transformer
Python 3
C++
VueJS
Azure
OpenCV
Image segmentation

Visual search for wood patterns across a catalog

The problem: An architecture firm needed to identify wood patterns in design references and surface similar in-stock patterns from its catalog, a manual task that took designers hours per project.
What we built: An image-segmentation pipeline isolates wood-pattern regions, Vision Transformer embeddings plus structural and histogram comparison drive a similarity index, and a gallery front end handles upload, query, and curated results over an indexed in-stock catalog.

45%

time saved on pattern search

95%

match accuracy vs. manual baseline

35%

less material waste

200%

catalog growth indexed in six months

6 members, 8 months, 9 features.

Further engagements

Where the numbers stay with the client

Regulated and internal systems we can describe, but whose results belong to the organizations that run them.

Industrial equipment service

Confidential

A generator-manual assistant that lives in MS Teams

Field technicians needed faster, more accurate answers and parts lookups from generator manuals without leaving Microsoft Teams, combining semantic search with structured parts data and returning grounded, auditable answers. An agentic Hybrid RAG assistant inside Teams. An ingestion pipeline builds a page-level vector index alongside normalized SQL parts tables; an orchestration layer routes each query between semantic RAG and SQL tools, and answers return with source-linked citations.

Manual lookups replaced by conversational queries, every answer grounded in source documents with required citations [1][2][3], and no new tool to learn since the assistant joins existing chat. 4 members, 6 months.

Education

Vietnam

Grading cycles cut from a week to hours

Grading ran on week-long cycles, and teachers had no consolidated view of how each learner was progressing. Edu Insight pairs a GPT-4-turbo teaching assistant with a 25-feature regression model that scores seven skill areas, feeding an LMS analytics dashboard. Built on Azure with SQL Server and a Vue front end.

Recruitment

Vietnam

Interview evaluation from two hours to fifteen minutes

More than 150 technical interviews a month, scored by different evaluators against inconsistent standards. Faster-Whisper transcription feeds GPT-4 rubric analysis that produces an automated PDF scorecard per interview. Runs on AKS with Prometheus and Grafana observability.

Recruitment

Hong Kong

Seven thousand CVs a month, shortlisted in minutes

Inbound CVs grew from five to seven thousand a month; manual ranking could not keep pace with open roles. A semantic matching pipeline: Apache Tika ingestion, OpenAI embeddings indexed in Pinecone, and a custom scoring API on FastAPI, deployed to EKS.

Recruiters report higher candidate-fit quality since launch.

Investment banking

Hong Kong

Multi-agent drafting for IPO documentation

IPO filings consumed hundreds of hours per document under strict HKEX and Nasdaq compliance requirements. DraftIPO runs ingestion, compliance, generation, and validation agents on CrewAI and LangChain, with a GPT-4o, Gemini 1.5, and Llama 3.1 ensemble over a Qdrant vector store, on Kubernetes.

Automates drafting across multiple exchanges. Results are confidential to the client.

Put your workflow on this page

Tell us where the work piles up. We scope the first engagement around the outcome, and we agree on the number that defines done before we build.

Get in touch Explore the services