Work
AI that made it to production.Measured in business numbers.
Every engagement here shipped: models running on client hardware, workflows their teams use daily, systems still in production.
Featured case
200K medical scans a month,
moved off the cloud.
A Swiss hospital group ran diagnostic imaging on AWS. The models now run on their own GPUs at the same throughput, for 95% less infrastructure cost.
−95%infrastructure cost
200Kscans processed monthly
$5.4Msaved over three years
PyTorch · CNN · NVIDIA GPUs · PACS · On-premRead the full case
All work
Every project we have
been cleared to publish.
Outcomes as infrastructure cost, throughput and adoption, not model scores. Where a client has not cleared a number, the card says what we built and nothing more.
01
Private hospital group, Switzerland
Diagnostic imaging AI moved off AWS onto their own hardware, 200K scans a month at the same throughput
−95% infra cost
02
Engineering consultancy, Germany
Lumen deployed as the company-wide AI workspace on four L40S GPUs, replacing EUR 14K/month of cloud subscriptions
94% monthly active
03
International law firm
A 72B model reading depositions and contracts on the firm's own H100s, so privilege never leaves the building
72B on-prem
04
Electricity distributor, Europe
A 72B model serving 28K operational documents fully offline inside the OT boundary, updated only through signed packages
28K docs, air-gapped
05
Document AI SaaS, Europe
Eval-gated migration from cloud LLM APIs to fine-tuned open-weight models on the client's own GPUs, at measured quality parity
−75% inference cost
06
Commercial lender, US
Field-level loan document extraction inside the lender's network, behind deterministic validation and a replayable audit trail
73% fields auto-posted
07
Specialty medical group, Chile
Referral and discharge-summary extraction where patient records never leave the network, built ahead of Chile’s Ley 21.719
45K documents monthly
08
P&C insurance carrier, Europe
Classify, extract, route pipeline for 16K claim bundles a month inside the carrier's network, with confidence-gated review
63% auto-routed
09
Wholesale distribution group
A Neo4j knowledge graph over the supply chain, answering multi-hop questions plain RAG cannot reach
12M-node graph
10
Wholesale distribution group
Fourteen scattered systems consolidated into one AI-ready layer, with query routing across CAG, RAG, and GraphRAG
14 sources unified
11
Electromechanical manufacturer, Vietnam
Vision-based quality inspection and predictive maintenance on the production line, built for factory constraints
99.2% defect detection
12
Liquid-bulk forwarder, Latvia
Two-week diagnostic across freight operations and back office, with an honest build-versus-skip verdict on each opportunity
7 opportunities ranked
13
Construction group, Europe
Face-recognition attendance across active sites, then a data layer and roadmap scoped to what field teams would actually use
500+ workers tracked
14
Agricultural trading group, China
Twenty years of trade data pulled into the CRM, with vision AI checking export documents and agents drafting deals
−90% manual entry
15
Grocery retail chain, US
A five-stage rollout from data infrastructure to demand forecasting, now moving into shelf-level computer vision
2 of 5 stages live
16
Conversation-analytics startup
WhatsApp and Telegram threads turned into tasks and risk flags, from a stalled MVP to a RAG system in production
6-week MVP
17
B2B outreach agency
Telegram outreach automated end to end, from lead research to message drafting, with human review before every send
36.4% engagement
18
Multilingual YouTube network
A multi-agent pipeline reading channel performance across three languages and returning weekly content briefs
50+ channels analyzedStart a project
“Every project on this page started with the same call: what you have, what it costs you, and whether AI actually fixes it. Sometimes the honest answer is no. When it is yes, this page is what the work looks like a year later.”
Sebastian Mondragon, Founder of Particula Tech