Selected work

Private AI systems, built around real constraints.

Selected work across private AI implementation, workflow automation, education technology, and synthetic data.

Dashboard showing recurring IT ticket clusters by root cause

Public sector & enterprise

IT Service Management Automation.

In collaboration with DigitFlow. Now available as a product package.

Problem
A chaotic backlog of mixed-media support tickets made recurring problems, root causes, and operational priorities hard to see.
Constraint
The workflow involved internal operational data and had to strictly respect data control requirements.
System
Built a secure local open-source AI pipeline using OCR and LLMs to digest attachments, summarize ticket content, cluster issues, and surface root causes through a dashboard.
Result
Manual ticket triage and root-cause discovery were replaced by an automated clustering pipeline running inside controlled infrastructure, so IT teams could act on cluster-level patterns instead of reviewing tickets one by one.
Education platform interface supporting student writing feedback

Education tech

AI for Education.

Backend and AI infrastructure for Mark My Words, supporting real-time student feedback.

Problem
Educators needed to provide scalable, personalized feedback on student writing without being buried in repetitive grading work.
Constraint
AI could support the workflow, but educator judgment and review needed to remain central to the learning process.
System
Collaborated with Mark My Words to build backend infrastructure integrating RAG models, fine-tuned LLMs, and asynchronous pipelines for classroom-scale traffic.
Result
Personalized feedback on student writing now moves through an automated pipeline with explicit educator review before final delivery.
Synthetic dataset visualization for privacy-preserving model validation

Privacy tech

Synthetic Data Engine.

Privacy-first model validation for teams that need realistic data behavior without exposing real records.

Problem
Models and pipelines needed to be validated, but using real data was impossible because of privacy constraints, and toy data was not realistic enough.
Constraint
The generated data had to preserve useful distributions, edge cases, and behavior patterns without exposing sensitive records.
System
Conducted statistical and logical analysis of real data structures, then recreated high-fidelity synthetic datasets that preserved useful distributions, edge cases, and user behavior patterns.
Result
Synthetic datasets replaced production records in pre-production testing while preserving realistic structures, distributions, and edge cases.

Next step

Have a workflow like this?

If your company wants AI productivity without giving up control over data, internal knowledge, and business processes, that is usually the right place to start.

Book a Private AI Workflow Assessment Send a message