Project Shimmer is a multi-agent system for reviewing regulated documents. It checks documents against operator-supplied conventions and reference material, then proposes cited amendments for a person to accept or reject. The source documents remain unchanged.
The system combines model reasoning with typed findings, deterministic comparisons and cross-model verification. An append-only constitution governs changes and operator decisions. Each run retains its evidence: findings, citations, agent activity and an audit trail. It runs on local open-weight models or cloud backends, in English and Turkish.
A causal-inference-based extension of Shimmer measures whether a frontier model can audit the causal integrity of an observational research design: whether it catches the silent failures a trained methodologist would catch, such as a weak instrument, an absent diagnostic, or a method asserted outside its regime. Two decisions carry the design. Ground truth exists by construction, because known violations are injected into synthetic designs before the model sees anything. And the setup is de-primed: the model receives raw, unlabeled material and no checklist, because a checklist measures recall, not judgment. On its first run the system recovered a deliberately flawed design and flagged a planted weak instrument from raw data.
The Research Cloud Platform is a Linux-based research-computing and data platform. It combines Spark processing, FastAPI serving, Docker, local Kubernetes with ingress and security controls, Terraform, CI/CD pipelines that publish container images, and monitoring with Prometheus and Grafana. Additional layers cover governance, lineage and cost-aware workload placement.
Its MLOps workload is a waste-demand and municipal-operations simulator for Karşıyaka. The workflow spans training and experiment tracking, batch and online inference, drift monitoring and retraining decisions.
A local LLM agent already handles tool calling, policy checks, human approval and auditing. It is being expanded into a full agent that can run the platform's own capabilities.
The Agricultural Documents Terminology and Unit Wizard reads agricultural reports in PDF, Word, Excel, CSV, JSON and plain-text formats, in English, Turkish and Portuguese. A trained multilingual XLM-R model interprets each measurement column. Values are then converted deterministically to canonical units, with full provenance retained for every number.
Explicit source units take precedence over model predictions. Ambiguous units are referred for human review rather than guessed, and approved decisions are retained so the same question is not asked twice. A single engine powers a desktop application, a versioned REST API and a command-line tool, all backed by a regression suite.
Values Pipeline, the LSE dissertation, classified approximately 25,700 UN General Debate sentences from 61 countries against ten World Values Survey dimensions, using SBERT embeddings and CatBoost classifiers with human-validated labels and a permutation-tested divergence index. Awarded a distinction; a journal article is in preparation.
Cosmic Horror Index: an NLP pipeline scoring thirteen cosmological traditions across ten weighted axes, written up in a Medium essay. A replication study that stress-tested a published instrumental-variable paper and showed its instrument weakening under unreported transformations. A survey-attrition classifier built on gradient boosting and neural networks. The rest is on GitHub.