Ömer Faruk Koç
Building production AI/ML platforms, agentic systems, retrieval infrastructure and reliable data systems.
3+ years building and operating production ML, data and GenAI systems, with current work focused on agent reliability, evaluation, observability and platform control.
Request lifecycle: open a stage
Trace + Metrics
Public engineering projects built around control boundaries, reliability, evaluation, observability and explicit trade-offs.
AI Infrastructure / Platform Engineering
A self-hosted Kubernetes ML platform for scheduled scoring, versioned datasets, model monitoring and governed serving, backed by a PostgreSQL control plane and a separate Go inference gateway.
Scheduled scoring and quality monitoring
1M ROWS
Real NYC TLC data, local acceptance, 8 Oct 2026
One million February 2025 taxi trips scored through a scheduled Argo pipeline and matched to ground truth with 100% coverage.
AI Reliability / Execution Infrastructure
Customer-service agent platform where the LLM proposes refunds, cancellations, lookups, tickets and escalations while deterministic software owns scope, policy, confirmation, revalidation, idempotency and execution.
Execution authority
SERVER-OWNED
LLM proposal → deterministic execution
Authentication, scope, policy, confirmation, revalidation, idempotency and business execution stay outside the model. The exercised D2c slice recorded 0 unauthorized mutations and 0 unsafe executions.
Agent Systems / AI Platform
Evidence-driven Kubernetes incident investigation: an in-cluster read-only Connector, deterministic causal engine, replayable evidence and an interactive operator console.
Held-out testbed
Cause named in 27/27 runs
HOLDOUT23, engine 2.17.0, 0 model calls
Nine fault variants, three fresh seeds each: 27/27 valid runs, the true cause and instance named in every run, zero false RESOLVED, zero false strong authority and the decoy never named.
Generative AI / RAG Platform
Agentic document research workspace: PydanticAI agents on an application-owned typed research graph search authorized sources, read original PDF pages, extract values and calculate in code, then release answers with citations you can open.
Agentic question coverage
103/120 accepted outcomes
120 distinct long-document questions
81 source-correct answers, 4 qualified calculations and 18 correct refusals, each question counted once and checked against the original PDF.
MLOps / AI Platform
Policy-driven ML release control plane combining progressive canary delivery, delayed-ground-truth quality gates, automated promotion and rollback, and desired-vs-observed routing reconciliation.
Routing control loop
DESIRED ↔ OBSERVED
Durable database state → router reconciliation
The database owns desired traffic; the router is restart-losable observed state. A worker-triggered reconcile tick repairs drift after a restart or failed push.
Structured Data / Governed Text-to-SQL
A governed Text-to-SQL workspace: ask a database in plain language, get a table and a chart with the SQL that ran, and build reports, while deterministic checks decide whether any SQL runs.
BIRD Mini-Dev 2026
76 / 100
PostgreSQL, a fixed random 100 of 500
Execution accuracy with gpt-6-luna on a fixed sample, 93 of the 100 questions challenging. The public leaderboard lists the same model at 51.8 on the full 500.
Distributed Systems / Streaming
Production-oriented event-driven commerce platform where Kafka may redeliver, but layered idempotency, transactional persistence and bounded failure handling protect durable business effects.
Sustainable isolated capacity
Workload-dependent
~1,050 evt/s at 42.8% fraud-eligible, ~1,600 at 0%
The processor has no single workload-independent ceiling, because fraud-eligible events take a longer path. The ~750 to ~1,050 evt/s (+40%) optimisation was measured on the 42.8% workload. Measured on an isolated local benchmark.
Capabilities are grouped by engineering evidence—not percentages or keyword counts.
Model lifecycle, progressive delivery, delayed quality feedback, policy-driven release control and observable production operations.
Retrieval, reranking, citation integrity, evaluation and private open-source model serving.
Event processing with explicit delivery guarantees, failure handling and measured service limits.
Batch and near-real-time pipelines, transformation systems, quality controls and data lineage.
Stateful agent workflows with deterministic control boundaries, confirmation and recovery, secure tool interfaces, evaluation and observability.
Fibabanka, Analytics Center of Excellence
Built and operated production ML, data and Generative AI platform capabilities across analytics workloads.
Professional experience informs the engineering questions explored in public projects; the public repositories are independent work.
Short technical pieces on reliability, evaluation, retrieval, distributed systems and production AI.
Making agentic development verifiable rather than merely autonomous: external gates decide whether work is acceptable, and hashed baselines stop the agent from redefining what acceptable means.
A practical execution model for tool-using agents built from typed proposals, deterministic policy, durable confirmation, revalidation, idempotency, and audit.
A cross-lingual reranker moved Recall@5 from 0.9563 to 1.0000. The aggregate is the least interesting number in that sentence, and the latency figure means less than it looks like.
402 source-grounded nodes and 538 validated relationships across experience, projects, technologies, concepts, evidence and current learning directions.
Open to remote opportunities, UTC+3 / Istanbul
Available for new opportunities
