Simas Razinskas · AI developer & consultant
Currently: turning AI workflows, agent systems and internal automations into software teams can rely on
Twelve production systems, written up the way they were built: the problem, the constraint that shaped the design, and how the system behaves when things fail. Clients stay anonymous. The engineering doesn't.
A stateful app needed frequent releases without downtime, broken containers reaching traffic, or bad migrations destroying the rollback path. Only a tested image tag can be promoted, traffic moves after health passes, and a failed release rolls back before users see it.
Read the case study →AI agents needed to work from the same queue as humans without double-claiming tasks or hiding failures. Agents can claim, heartbeat, complete, fail, and escalate work safely while humans watch the same board update in real time.
Read the case study →Third-party webhooks retried, duplicated, and arrived out of order; silent loss would only surface later as broken reports. Data-loss risk moved from a quarterly investigation to continuous checks that detect gaps, repair them, and expose operational health.
Read the case study →Most of this shipped under NDA — so it's described by what it does, with public work to back it up.
End-to-end translation and adaptation of video, audio and image content — transcription, machine translation, voice synthesis and on-image text replacement, wired together to run at scale.
Monorepos stitching dashboards, queues, APIs and third-party integrations into the daily operational workflow of marketing and ops teams.
Retrieval-augmented helpdesk assistants, automated ticket triage and internal copilots — LLMs put to work against real business data and processes.
Figma plugins and Adobe scripting for design automation, plus Chrome extensions that fold internal tooling directly into the browser.
Customer and credit management systems, storefronts and learning platforms — typed full-stack apps on Next.js with Postgres behind them.
Interactive map products over OpenLayers and GeoServer, including EV-charging location and routing data through open standards.
Cross-platform apps with React Native / Expo, native mobile work, and a native macOS utility in Swift.
CLIs and self-hosted delivery pipelines — including the agent-operated platform that runs this very domain.
Twelve case studies, anonymized down to the engineering: what broke, what constrained the fix, and what shipped. AI pipelines, distributed systems, infrastructure.
Build-once / promote-only images, staging-tested with live migrations, then swapped atomically behind a reverse proxy. A container health gate decides each release: pass drains the old color, fail rolls back automatically with the old version still serving. Pre-migration database snapshots and expand/contract migrations make even schema changes reversible.
Docker Compose · Traefik · Postgres · shell orchestrationA distributed kanban where autonomous agents claim work via FOR UPDATE SKIP LOCKED under renewable leases, with a lazy TTL sweeper (no background worker), three-strike escalation to a human, and an append-only event log. Boards sync in real time over Postgres LISTEN/NOTIFY fanned out as server-sent events.
Drizzle · Postgres LISTEN/NOTIFY · SSE · contract-first APIA durable Postgres-backed queue accepting third-party webhooks with at-most-once delivery via upserts keyed on external IDs. Three self-healing layers guard against silent data loss: dual-trigger reconciliation, missing-ID backfill detection with tombstone anti-joins, and 30+ automated anomaly checks on latency and orphaned records.
Go · Postgres (SKIP LOCKED, matviews) · HMAC · advisory locksReplaced a stack of brittle no-code automations with a hand-rolled durable queue: lease columns, claim counts and heartbeats dispatching rate-limited jobs to external APIs, child idempotency keys for mid-flight recovery, and shadow / dry-run modes for testing safely in parallel with production.
Go · Postgres leases · gocron · bearer + HMAC authAn end-to-end dubbing pipeline: transcription with speaker diarization and word-level forced alignment, neural voice cloning, and per-segment audio time-stretching so synthesized speech matches the original timing without drift across multi-minute media — with vocal/instrumental separation to keep backing audio clean.
WhisperX · XTTS voice cloning · Demucs · FFmpeg · CTranslate2Extracting and re-injecting text across 3,000+ segments inside packaged document formats (nested ZIP/XML trees) while preserving exact element indices, inline formatting and layout — surviving a full export → translate → reimport round trip without corrupting structure.
Python XML/ZIP tooling · multi-format CAT interchangeDiscovering better translation prompts with a genetic algorithm — elitism, adaptive per-mutation weights, crossover and diversity maintenance — scored by a learned semantic-similarity metric under a multi-objective fitness function balancing quality, length and vocabulary.
BLEURT · LLM APIs · evolutionary operators · NumPyDriving multi-minute image-generation workflows on serverless GPUs over WebSockets: enqueue, poll, stream large binary outputs, and recover job state across mid-inference network drops with bounded exponential backoff and clean worker teardown to prevent VRAM leaks.
RunPod · ComfyUI · WebSockets · S3 offloadRouting generation requests across many model providers behind one uniform adapter interface, with per-model concurrency limits, automatic failover, and session / daily / total cost aggregation — exposed over both async REST and streaming gRPC.
Go adapters · gRPC / protobuf · persisted in-memory queueA type-safe full-stack platform enforcing tenant isolation on every query and a five-role permission hierarchy at the API boundary, with GDPR audit trails, multi-currency money handling and approval workflows.
tRPC · Prisma · Postgres · Next.jsSub-3-second polling that detects new trades, sizes them proportionally to a different account's balance, rejects on excessive slippage, and settles against an on-chain order book — deduplicating by transaction hash while bounding memory and surviving connection-refresh cycles.
Python · Web3 RPC · on-chain CLOBLocal-first voice-to-text: low-level audio capture → on-device speech recognition → context injection from screen OCR, clipboard and selection → optional LLM cleanup → accessibility-API paste — orchestrating five macOS permissions and mid-recording device hot-swap with non-fatal failure handling.
Swift / SwiftUI · whisper.cpp · Accessibility & CoreGraphics APIs2026
This site: Next.js 16 with Cache Components, statically generated and continuously deployed straight from CI.
2026
A single VM running multiple isolated app stacks behind Traefik with Let's Encrypt, full Prometheus/Grafana/Loki observability, alerting to ntfy, and scripted backups.
2025 →
Building software primarily with AI coding agents on top of strict typing, boundary validation and deterministic verification.
2024 → 2025
End-to-end localization of video, audio and image content — speech-to-text with diarization, neural voice cloning, frame-accurate audio sync, and GPU inference orchestrated on serverless hardware.
2024 →
Go and TypeScript monorepos replacing brittle no-code automations: durable database-backed job queues, idempotent webhook ingestion, agentic helpdesk assistants, and operations dashboards with row-level multi-tenant isolation.
2023 → 2024
Production CRMs, e-learning platforms, e-commerce and image-generation gateways — type-safe end to end on Next.js, tRPC and Postgres, with the first wave of LLM features wired in.
2022 → 2023
Mapping products over GeoServer/OpenLayers with cross-projection transforms, real-time market-data and trading tooling, design-tool and browser plugins, and native macOS and mobile apps.
since 2014
Over a decade of shipping on GitHub — a long tail of languages, frameworks and side projects that predate the work above.
I use this domain as a lab — ideas only get interesting against a real production environment: real users, real deploys, real failure modes. Most of the code here is written by AI coding agents working a kanban board: I file cards with acceptance criteria, agents claim them through a leased API and work behind a strict verification gate — types, lint, tests, dependency boundaries — parking anything ambiguous as a question for me.
The infrastructure is deliberately self-hosted: a single VM behind Traefik, with continuous delivery from CI through a container registry, container health checks and self-hosted analytics. Owning the whole stack keeps the feedback honest — when something breaks, it's mine to understand.
Simas Razinskas · AI developer & consultant
Book a call and we'll map it out: what it should do, where it will break, and the smallest version worth building first.
The public edge of work that mostly ships private: current practice, visible source, small reviewable commits.
Most repositories are private — counts include private contributions.