01/03
Built for pressure
Deterministic workflows, clearer failure modes, and decisions that do not collapse once the system has real users and real stakes.
// initialising …
// session ok
- NAME
- Ali Uyar
- ROLE
- AI Systems Engineer
- LOCUS
- Vienna, Austria · UTC+01
- STATE
- ● In a full-time role
objective
I build AI products that hold up in practice —
retrieval, agents, and model-reliant workflows where
reliability, evidence, and technical judgment matter
more than the demo.
§ 01 Principles
02/03
Evidence before confidence
Evaluation, traceability, and auditability belong in the product itself, not as reassurance added later.
03/03
Judgment over theatre
Useful AI systems come from better sequencing, better constraints, and clearer choices — not just a more polished demo.
№ 01
Inside the Multilingual Transformer: Computational Neuroanatomy of Shared and Language-Specific Brain Alignment.pdf
Shows which internal transformer computations actually drive shared cross-lingual alignment, giving multilingual model developers a mechanistic target instead of a single scalar benchmark.
multilingual-nlpbrain-alignmentmechanistic-analysis
№ 02
Language-Neutral or Language-Specific? Disentangling Multilingual LLM Subspaces with Naturalistic fMRI.pdf
Provides a reproducible subspace decomposition that teams can reuse to audit whether multilingual encoders carry cross-lingual content structure versus language-specific residuals.
multilingual-nlpfmri-encodingrepresentation-analysis
№ 03
Overlay Algebra: Causal Composition of Operational Behaviors in Frozen Language Models.pdf
Lets teams ship and audit reusable output-contract adapters as residuals on a frozen base instead of re-training whole behaviors from scratch.
frozen-modelsadapter-residualsauditable-evaluation
№ 01
DecisionGraph.py
DecisionGraph is a local-first library for immutable AI decision traces. It is used for audit-ready records and deterministic replays in compliance and approval workflows.
Python source
№ 02
SchemaPilot.py
SchemaPilot is a governance-first data platform for building AI-ready datasets. It is used to control schema quality, access policies, and deterministic pipelines.
Python source
№ 03
№ 04
cost-watchdog.ts
Cost Watchdog is a self-hosted cost monitoring and anomaly detection platform. It is used to ingest cost data, detect anomalies automatically, and notify teams before overspend appears in monthly or yearly reporting.
TypeScript source
№ 05
VeilPack.rs
VeilPack is an offline, fail-closed privacy gate for enterprise data pipelines. It is used to detect sensitive values, redact outputs, and ship verifiable data packs.
Rust source
§ 04 Outcomes
case_01
Policy-Enforced Context Runtime
- problem
- Plain RAG pipelines left governance gaps around policy, provenance, and deterministic failure handling.
- build
- Built a two-plane runtime with a non-privileged controller, policy-enforcing gateway, immutable evidence units, and deterministic terminal modes.
- result
- Delivered replayable, auditable AI retrieval runs with clear permission and evidence outcomes for safer production deployment.
case_02
SchemaPilot
- problem
- Enterprise data lived across inconsistent sources, making AI usage hard to govern, trust, and operationalize.
- build
- Built a governance-first data platform with gateway-enforced access controls, deterministic bronze/silver/gold pipelines, and operator-grade tooling.
- result
- Created an AI-ready, queryable data foundation with stronger auditability, safer defaults, and more predictable delivery workflows.
case_03
ContextGraph
- problem
- Teams lacked trustworthy process visibility across Slack, Jira, and GitHub without creating surveillance risk.
- build
- Built a self-hosted process intelligence platform with fail-closed permissions, private personal timelines, and k-anonymous analytics.
- result
- Improved process insight and next-step guidance while preserving privacy boundaries and policy-first controls.
§ 05 Focus
// past the demo, into the weather
When a team is past the demo stage and trying to make something dependable — what the system should be trusted to do, where evidence should exist, and which trade-offs are worth making.
That tends to sit between engineering, product, and delivery. I like small teams, clear questions, and work that leaves a system easier to understand than it was before.