ACAndrei Cruceru · AI Systems Engineer
Case Studies

Enterprise AI systems, in production.

A selection of anonymised engagements spanning data quality, legal automation, and transformer-based classification at scale.

dqp.engine · execution trace
live
step 01
Parse workflow
graph → DAG
step 02
Plan columns
column-aware
step 03
Check cache
content-hash keys
step 04
Execute tools
300+ DQ ops
step 05
Merge partial results
streaming
step 06
Emit insights
issues · fixes
▸ engine.run(workflow)
resolved columns=42 · cached 17/42 · executed 25 ops · async · throughput ~2.4k rows/s
01Case Study

DQP: Enterprise Data Quality Platform

Column-aware, cache-aware execution engine powering 300+ data quality tools, agents, and MCP tooling.

300+DQ tools
1000s/sRows / sec
AsyncGraph exec
PythonMLAI AgentsMCPAsync
Read case study
legal.intake · triage pipeline
days → minutes
  1. stage 01auto
    Inbound enquiry
    email · form · portal
  2. stage 02auto
    Relevance filter
    ML classifier
  3. stage 03auto
    Case classification
    department + case type
  4. stage 04auto
    Cost & time forecast
    regression models
  5. stage 05auto
    Legal retrieval
    UK laws · case law
  6. stage 06auto
    Fee-earner recommendation
    ranking model
  7. stage 07HITL
    Department head review
    human-in-the-loop
  8. stage 08auto
    Assignment + client update
    structured record
Custom ML inference
Email filter · case classifier · cost/time regression · fee-earner ranker
Retrieval layer
UK laws + court cases indexed for contextual case support at assignment.
Automation surface
Follow-ups · structured records · department alerts · client updates.
02Case Study

Legal Intake, Triage & Assignment

End-to-end onboarding pipeline: filter, classify, forecast, retrieve, recommend, and assign fee-earners.

+50%Onboarding cap.
£300k+Yr-1 saving
Days→MinTriage time
Custom MLRAGHITLWorkflow
Read case study
classification.pipeline · inference path
2M+ / month
01
Product descriptions
unstructured text
02
Preprocessing
clean · normalise · tokenise
03
Transformer model
fine-tuned encoder
04
ONNX Runtime
optimised inference
05
.NET API
enterprise integration
06
Category output
600+ classes
▸ transformer.predict(description) → category · confidence 0.97 · latency ~12ms · runtime onnx
03Case Study

Product Classification System

Transformer classifier for a global maritime client, deployed on ONNX Runtime behind a .NET API.

2M+/moPredictions
600+Categories
95%+ F1Top 85%
TransformersNLPONNX.NET
Read case study