ÖFK
Menu

Public engineering work

Projects

Evidence-led systems work across AI platforms, agent systems, retrieval, data infrastructure and distributed systems—with reliability, evaluation, observability and trade-offs made explicit.

View as Markdown

Featured work

Seven current projects selected for depth, engineering evidence and relevance to AI platform work.

Project / 01current

AI Infrastructure / Platform Engineering

ML Platform

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.

  • Python
  • Go
  • Kubernetes
  • KServe
  • FastAPI

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. Historical-data replay on one ARM64 kind node with warm images, not production capacity.

View case study GitHub for ML Platform (opens in a new tab)
Project / 02current

AI Reliability / Execution Infrastructure

Agentic Customer Service Platform

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.

  • Python
  • FastAPI
  • LangGraph
  • SQLAlchemy
  • PostgreSQL

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.

View case study GitHub for Agentic Customer Service Platform (opens in a new tab)
Project / 03current

Agent Systems / AI Platform

Agentic SRE

Evidence-driven Kubernetes incident investigation: an in-cluster read-only Connector, deterministic causal engine, replayable evidence and an interactive operator console.

  • Python
  • Kubernetes
  • LangGraph
  • gRPC
  • FastAPI

Evidence-driven investigation

DETERMINISTIC

Read-only Connector · causal engine · operator console

Nineteenth HOLDOUT, engine 2.8.0 (2026-10-11): 24/24 valid lab runs, 12/99 scored diagnoses RESOLVED, zero false RESOLVED and zero false strong authority. Every resolved diagnosis names its run’s true experiment. These are controlled testbed results, not production evidence.

View case study GitHub for Agentic SRE (opens in a new tab)
Project / 04current

Generative AI / RAG Platform

Knowledge Base RAG

Local-first multilingual RAG platform where tenant scope, evidence construction, support-unit identity and occurrence-aware validation are separate boundaries, and where a change is adopted only if it passes a decision rule frozen before the result was known.

  • Python
  • FastAPI
  • React
  • TypeScript
  • Vite

Multilingual reranker benchmark

MRR 0.367 → 0.956

220 questions · frozen multilingual benchmark

The multilingual reranker improved Cross MRR from the prior English reranker's 0.367 to 0.956 and reached Recall@5 of 1.000 on both TR-to-EN and EN-to-TR slices. These are retrieval/ranking metrics, not final-answer accuracy.

View case study GitHub for Knowledge Base RAG (opens in a new tab)
Project / 05current

MLOps / AI Platform

ModelOps Control Plane

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.

  • Python
  • FastAPI
  • SQLAlchemy
  • SQLite
  • Alembic

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.

View case study GitHub for ModelOps Control Plane (opens in a new tab)
Project / 06current

Structured Data / Governed Text-to-SQL

DecisionSQL

Governed one-shot Text-to-SQL for enterprise analytics: the model emits one typed decision, deterministic software owns SQL admission and read-only execution, and correctness is measured by execution against counterfactual database states.

  • Python
  • FastAPI
  • PostgreSQL
  • SQLAlchemy
  • Alembic

Execution authority

SERVER-OWNED

Model proposes SQL · deterministic admission decides

An ANSWER + SQL submission crosses parse, global policy, request-scoped relation authority, grain safety, EXPLAIN and a cost gate, and only an accepted immutable QueryPlan reaches the restricted read-only executor. There is no raw unsafe fallback, and the executor never runs SQL the safety service did not admit.

View case study GitHub for DecisionSQL (opens in a new tab)
Project / 07current

Distributed Systems / Streaming

Real-Time Commerce Platform

Production-oriented event-driven commerce platform where Kafka may redeliver, but layered idempotency, transactional persistence and bounded failure handling protect durable business effects.

  • Python
  • TypeScript
  • FastAPI
  • Next.js
  • Kafka

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. Local isolated benchmark, not production capacity.

View case study GitHub for Real-Time Commerce Platform (opens in a new tab)

Supporting / evolution

Supporting work

Focused tools and earlier foundations that show architectural iteration and how narrower systems evolved into broader platforms.

Supporting workcurrent

Model Training / Fine-Tuning

CauseTune

LLM fine-tuning laboratory built around the full craft, not one training run: capability-gap measurement, QLoRA recipe selection under a hardware and data budget, held-out generalization testing, failure-mode and calibration analysis, and cross-model replication.

  • Python
  • PyTorch
  • Transformers
  • PEFT

Measured specialization gain

+34.03 pp diagnosis exact match

Frozen 144-case synthetic benchmark

Untouched Qwen3-4B scored 65.28% on a benchmark frozen before training and excluded from checkpoint selection; the tuned adapter reached 99.31%. A second, independent blind challenge the adapter never saw later confirmed the gain held at 98.33%, while a cheaper alternative recipe did not hold as well on the same challenge. 99.31% is benchmark accuracy on synthetic cases, not production accuracy.

View case study GitHub for CauseTune (opens in a new tab)
Supporting workcurrent

Agent Infrastructure / Developer Tooling

Repo Context Forge

Local-first MCP repository intelligence and agent platform for deterministic, source-grounded code analysis within strict read-only boundaries.

  • Python
  • MCP
  • FastMCP
  • Python AST

Configured MCP surface

40 tools

Six local MCP servers

The platform configures 40 tools across six servers; the read-only local agent uses a bounded subset by default.

View case study GitHub for Repo Context Forge (opens in a new tab)
Supporting workcurrent

Data Engineering / Lineage

dbt Feature Lineage

Local-first dbt analysis for model dependencies, cross-model column lineage, change impact and query flow without a live warehouse.

  • Python
  • dbt Core
  • FastAPI
  • Next.js

Analysis scope

Cross-model column lineage

Manifest-aware + static analysis

Traces columns upstream and downstream and summarizes direct versus transitive downstream impact without a live warehouse connection.

View case study GitHub for dbt Feature Lineage (opens in a new tab)
Supporting workcurrent

Infrastructure as Code / Platform Engineering

Terraform Docker Infrastructure Lab

Modular Terraform infrastructure lab with state-safe refactoring, native IaC tests, CI security gates, observability, and end-to-end failure validation.

  • Terraform
  • Docker
  • FastAPI
  • PostgreSQL

State-safe module refactor

0 add · 0 change · 0 destroy

Verified Terraform migration

The original root resources were reorganized into network, application and observability modules through explicit Terraform moved blocks while preserving managed resource identities and avoiding infrastructure recreation.

View case study GitHub for Terraform Docker Infrastructure Lab (opens in a new tab)
Supporting workCompleted series

Predictable backend behavior under load and failure.

Go Resilience Labs

A focused series of six Go labs exploring how backend systems remain predictable under load, concurrency, and partial failure — covering latency measurement, idempotency, admission control, dependency resilience, backpressure, and end-to-end deadline budgeting.

  1. Measureapi-prober

    Bounded-concurrency HTTP probing with throughput measurement, latency percentiles, status distributions, connection reuse, graceful cancellation, and race-tested concurrency.

    GitHub for api-prober (opens in a new tab)
  2. Correctnessidempotency

    Concurrent idempotent request handling backed by PostgreSQL transactions and UNIQUE constraints, preventing duplicate side effects while safely replaying completed results.

    GitHub for idempotency (opens in a new tab)
  3. Admission Controlrate-limiter

    Local token-bucket and Redis-backed distributed rate limiting with atomic decisions, explicit rate-limit responses, and concurrency-tested global limits.

    GitHub for rate-limiter (opens in a new tab)
  4. Resilienceretry-circuit-breaker

    Dependency-free bounded retries with exponential backoff and jitter, retryable HTTP classification, response-body lifecycle handling, and a concurrency-safe circuit breaker.

    GitHub for retry-circuit-breaker (opens in a new tab)
  5. Overload Controlbackpressure

    Bounded queues and fixed workers demonstrating saturation behavior, enqueue timeouts, explicit overload rejection, context cancellation, stats, and graceful draining.

    GitHub for backpressure (opens in a new tab)
  6. Time Budgetdeadline-budget

    End-to-end deadline propagation with child timeout budgeting, safety reserves, fail-fast budget exhaustion, budget-aware retries, and parent-child deadline guarantees.

    GitHub for deadline-budget (opens in a new tab)
GitHub for Go Resilience Labs (opens in a new tab)