sortInc. is an engineering studio in Seoul building knowledge graphs, RAG pipelines and AI agents — plus the gateways, APIs and operator interfaces that make them usable. Not demos: guardrail chains with audit trails, detectors with false-positive filters, test suites in the hundreds.
Capabilities
LiteLLM proxies with custom guardrail chains: PII masking, term blocking, high-impact-decision detection, emergency stop, and hash-chained JSONL audit evidence — mapped to Korea's AI Framework Act.
Retrieval that cites its sources. pgvector and RAGFlow pipelines alongside Neo4j + n10s ontology stores, so semantic memory is queryable as RDF instead of an opaque embedding blob.
Two-stage inference where a fast detector proposes and a VLM verifies — YOLO recall first, qwen3-vl precision second, so shadows, puddles and manhole covers stop being potholes.
End-to-end LoRA and QLoRA pipelines for local models: synthetic data generation, training, eval, adapter merge, GGUF conversion and re-registration into Ollama — text and vision, on your own hardware.
Real-time Omniverse/USD plant twins driven by a single versioned WebSocket contract, and vector-tile map stacks — tiles, Korean glyphs, sprites, viewer — that run with zero runtime network access.
The unglamorous half that decides whether AI ships: FastAPI services, React + Vite frontends, browser extensions, Postgres migrations, Docker Compose, nginx, and forecasting on real operational data.
Open source
Everything below is public on github.com/sort-tech. Clone it, run it, audit it. We publish the tooling we use ourselves, because the fastest way to judge an engineering partner is to read their commits rather than their case studies.
Turns a screenshot, scan or whiteboard photo into a real .pptx whose tables and text boxes are native, editable shapes — never a flattened image.
Step-by-step tutorial and working code for fine-tuning LLMs with LoRA, QLoRA and Unsloth / Hugging Face SFTTrainer.
A lightweight, embeddable LLM chat widget — add an AI assistant to any site with a single script tag. Side-panel mode, page summarisation, LiteLLM backend.
Ontology ingestion for Neo4j with the n10s semantic plugin and APOC — RDF in, a queryable knowledge graph out. The backbone of our graph-based agent memory.
An enterprise blockchain framework — distributed-ledger groundwork for auditable, tamper-evident records, the same instinct behind our hash-chained AI audit trails.
This site. One HTML file, no build step, no framework — served straight from GitHub Pages. The source of the page you're reading is one click away.
These are public forks we maintain for client deployments — customised builds, embeds and integrations of upstream projects. Listed for transparency about the platforms we operate day to day, with full credit to their original authors.
Selected work
Client and internal systems beyond the public repos. Descriptions stay at architecture level where the work is under NDA.
A six-stage guardrail chain in front of a LiteLLM proxy. Requests pass through evidence logging, governance gating, PII masking, high-impact detection and transparency injection before they reach a model — and back out again in reverse on the response.
Every success and failure lands in a hash-chained JSONL audit trail, so what was sent, masked or blocked is provable after the fact.
Potholes, rutting and cracks, detected in two stages. YOLO runs at a deliberately low confidence threshold for recall; a vision-language model then interrogates each candidate crop for precision.
The response separates verified damage from rejected false positives with reasons, so operators can see why a shadow or tyre mark was dropped instead of trusting a single score.
A 2-bed 3-tower ultrapure water plant modelled in USD and driven live inside NVIDIA Omniverse. Resin bed heights, colours, flow direction and alarm lamps re-bind every frame from process data over one versioned WebSocket contract.
Built against a frozen wire specification with a mock process engine and a 940-plus test suite, so the customer's real engine drops in without touching the visual layer.
A full slippy-map stack for networks with no internet at all: OSM extracts clipped with osmium, built into MBTiles by tilemaker, served by tileserver-gl, with glyphs, sprites and MapLibre vendored locally.
One bootstrap script prepares, starts and verifies the stack across 30 automated checks, plus a React facility viewer for port infrastructure on top.
Agent memory stored as an ontology rather than a pile of vectors. Neo4j with the n10s semantic plugin and APOC holds RDF triples, so recall can be reasoned over, constrained and explained — and vector search becomes one retrieval path among several.
The ingestion layer is public as neo4j-n10s-pipeline.
A Chrome and Edge side-panel assistant grounded in the page you're reading. For long documents it selects only the passages relevant to the question, then streams an answer with evidence chips that highlight and scroll to the exact sentence.
Its embeddable sibling is public as llm-chat-widget.
Two parallel five-stage pipelines — a multimodal one turning receipt images into structured JSON, and a text one routing customer enquiries — sharing the same path from synthetic data through training, eval, adapter merge, GGUF conversion and Ollama re-registration.
The teaching version is public as llm-finetuning-tutorial.
Panel-structured forecasting with external covariates, honest baselines and rolling-origin backtests — including foundation time-series models next to classical ones, because a model only earns its place by beating a naive forecast on held-out windows.
A membrane-deionization sizing tool for water treatment engineers, reproducing the standard two-page feed/output projection sheet. The engine computes on real chemistry — ionic balance, Kohlrausch conductivity, Faraday ion transport and weak-acid dissociation equilibria.
Field-camera coastline imagery captured in a compact binary snapshot format, then scored by model to rank usable frames — with a self-contained HTML recording viewer for reviewing captures without installing anything.
Model Context Protocol servers and agent workspaces that give models real, bounded tools — plus visual flow orchestration with Flowise and RAG stacks wired to pgvector, so non-engineers can trace what an agent actually did.
Vision LLMs that emit schema-validated JSON instead of prose, then a deterministic builder that turns it into real artefacts — native PowerPoint shapes, structured receipts, routed tickets. The model proposes; the schema decides.
Open-sourced as image-to-ppt.
Stack
Gateways, local inference, tuning and evaluation.
Storage that can answer, and explain the answer.
Detection, verification and forecasting.
Typed contracts, tests, streaming.
Interfaces operators actually use.
Reproducible from a clean machine.
How we work
Most AI projects don't fail at the model. They fail because nobody wrote down what the interface promises, what happens when the model is wrong, or how anyone proves afterwards what it did. So we start there.
Talk through a projectWire schema, versioning, error codes and the boundary vocabulary get written and frozen before implementation. Both sides can then build in parallel against a mock.
Models mispredict. Pipelines get a verification stage, a rejected-with-reason channel and a human review path, so a wrong answer degrades instead of propagating silently.
Hash-chained event logs, policy versions and masking records. When a regulator, an auditor or a customer asks what the system did in March, the answer is a query, not an archaeology project.
Local models, vendored assets, Docker Compose, no runtime calls out. Plenty of our work is for environments where sending data to a hosted API is simply not an option.
Every repo states what was verified and what wasn't. A known-limitations section is worth more on handover than an optimistic README — and it's why our code is public.
About
sortInc. is an independent AI engineering practice based in Seoul, led by Jwp — an AI & backend engineer working on knowledge graphs, RAG pipelines and AI agents.
It exists because the interesting problems in applied AI sit exactly where most teams split responsibilities: between the model, the service around it, the operator's screen, and the compliance story.
We work across all four. A project can go from tuning a local vision model, to the FastAPI service that verifies its output, to the React dashboard an operator watches, to the audit trail the legal team needs — without three handovers and two integration projects.
Domains so far: port and maritime operations, water treatment and process plants, road infrastructure inspection, coastal monitoring, document automation and internal AI platforms.
Seoul, South Korea · remote-friendly
Open source by default. You keep the repo.
Send a short description of the problem and the constraints you're stuck with — on-prem only, no labelled data, a regulator asking questions. Those are usually the interesting part.
github.com/sort-tech · Seoul, Korea