Let's build AI apps with AI
Oridecon Labs is a software studio built around coding agents. A human sets the specs, agents like Claude Code do the typing, and it all runs on QuadKit — the Python framework we make for exactly that: open source and in-house.
THE THESIS
One developer, shipping at team speed
Agentic coding only pays off if what comes out stays consistent. QuadKit is what makes that true — the conventions are built into the framework rather than left to review, so an agent cannot drift from the pattern it was handed. Alignment buys speed. Speed buys the team.
- Human · decides Architecture, contracts, and what ships next. The judgement is never delegated — and with one developer, it never has to be.
- Agents · build The spec fans out across several agents at once because QuadKit is a pattern each agent can follow every time — a hierarchy it cannot violate, a type that forces the branch, a generator that writes the structure — not prose to memorize. Idea-to-running-code collapses to review time.
- QuadKit · aligns Every convention is enforced rather than suggested, so each file lands on the same shape as the last — whoever wrote it.
- App shipped · MVP
THE WORK
Ready to scale MVP
Built with Claude Code + QuadKit
The expensive part is already behind us. QuadKit — our building kit — is built and paid for, so funding goes into the product, not the platform. AI does the typing, the kit keeps every file on the same shape, and one developer holds the judgment — this build comes out fast, consistent, and at a fraction of what a team would cost.
Every product here started as an MVP.
Have one that needs building? That's the work we take on.
Gatewai · Dashboard
Codename: Gatewai
AI API Gateway
Every AI product rebuilds the same plumbing — provider keys, retries, quotas, cost attribution, billing. Gatewai ships it as one OpenAI-compatible endpoint, drop-in for any SDK: failover across 17 providers, per-model quotas, and a 10-stage relay that governs and meters every request. The wedge is zero-friction integration; the moat is the policy layer — channels, quotas, audit, settlement — that gets harder to rip out the more traffic it carries. Stripe subscriptions and top-ups are already built, and Cinestaq already routes its own production traffic through it.
What's built
- Byte-exact OpenAI-compatible API — any OpenAI SDK client works unchanged, so adoption starts with a base-URL change
- 10-stage relay pipeline — guards, governance, channel selection, protocol conversion, cost tracking, billing
- Circuit-breaker failover — 3 retry attempts across priority-weighted channels with health tracking
- Metered revenue — Stripe billing with pre-consume quota checks, post-relay settlement, subscriptions and top-ups
- React 19 operator console — channels, models, users, tokens, governance, playground, audit
What's next
- Public AI-API storefront — tiered subscriptions and per-request markup on relayed tokens
- Enterprise tier — dedicated instances, SLA, SSO and RBAC for larger contracts
- Multi-region deployment with cross-region failover — availability that can carry SLA-backed pricing
- Model fallback chains with quality scoring — better answers at the same spend
- Provider rate-limit autoscaling — grow throughput without renegotiating every provider deal
QK ERP · Good morning.
Codename: QK ERP
Marketplace-native ERP for HKTVmall merchants
Marketplace sellers in Hong Kong keep stock in a warehouse, orders on HKTVmall, and the truth in a spreadsheet — the incumbent answer is a year-long ERP rollout. QK ERP is the opposite: an MVP product merchants can run today, with 50 screens covering receive-to-settle — listing and order sync, picking and labels, invoicing, settlement. The moat is coverage: an open-source ERP addons parity ledger turns ERP depth into a shipping cadence, on an in-house stack of 10 QuadKit packages that keeps cost-to-serve flat as merchants grow — land per merchant, expand by module.
What's built
- MVP in production — 50 merchant screens covering receive-to-settle: listings, orders, stock, pick and pack, invoicing, settlement
- Marketplace-native by design — HKTVmall listing sync, order import, stock drift, returns and settlement as a first-class domain
- Multi-tenant SaaS foundation — tenant isolation, API-key scopes, ACL, audit trail and Prometheus metrics from day one
- Owned stack — 10 in-house QuadKit packages, no third-party per-seat fees, flat cost-to-serve as merchant count grows
- Audit-grade ledger — Decimal money, reviewable migrations, fail-secure tenancy: financials that stand up to diligence
What's next
- System of record — P&L, balance sheet, aged AR/AP and margin-by-order reporting to widen wallet share
- Full ERP coverage — close out the open-source ERP addons parity ledger, release by release
- Partner surface — view metadata, RPC and OpenAPI for integrations and a future app ecosystem
- Enterprise readiness — session admin, MFA and OAuth policy to move upmarket
Cinestaq · episode timeline
Codename: Cinestaq
AI Continuous Movie Series Production
AI video tools ship models; nobody ships finished episodes. Cinestaq is the production line: give it a script and a DAG workflow runs LLM planning, storyboard diffusion, video generation and neural TTS with cross-episode character and style consistency — the part every model demo skips. Every generation flows through Gatewai, so one gateway replaces a dozen provider SDKs: models swap the day prices drop, without a rewrite. The workflow engine is the product; episodes are the proof.
What's built
- Full AIGC pipeline — novel script to exported video episode via LLM + diffusion + video + TTS
- Cross-episode character consistency — characters persist across episodes with appearance tracking
- Style lock — visual style enforced per series for continuity
- DAG workflow engine — checkpoint/resume, conditional routing, saga compensation
- 17 media adaptors via Registry — swap image/video/TTS providers without code changes, so unit economics follow model pricing
- All generation delegated to Gatewai — one gateway, zero provider SDKs in the app, one meter for every model call
What's next
- Self-hosted inference — open-weight video, audio and diffusion engines to widen gross margin
- Marketplace for agent workflows and templates — recurring revenue on the engine
- Cloud rendering with usage-based billing — metered revenue per delivered episode
- White-label licensing for production studios — enterprise deals layered on usage
- API access for third-party integrations — platform revenue beyond the studio
Piccolina · Welcome
Pet owners Google symptoms at 2am and get conflicting advice from blogs, forums, and AI hallucinations. The real problem is information asymmetry — they don't know what they don't know. Piccolina closes that gap: describe symptoms or snap a photo, and a 6-stage RAG pipeline over a curated veterinary knowledge base answers in 30 seconds, not 3 days — urgency scoring, care plans, breed-specific warnings. The revenue ladder behind that daily-use hook is already mapped: per-pet premium, clinic SaaS across 673+ mapped PH clinics, telehealth and insurance referrals.
What's built
- 6-stage AI analysis pipeline — intake, parallel RAG retrieval, temporal health context, LLM analysis, enrichment
- Multi-query RAG — symptoms + breed predispositions + urgency + patient history, deduplicated and scored
- PostGIS vet discovery — 673+ Philippine clinics sorted by distance and emergency capability
- Community feed — location-aware posts with AI wellness summaries, rescue alerts, smart matching
- Temporal health context — longitudinal pet health tracking with activity logs, milestones, wellness scores
What's next
- Freemium launch — basic analysis free, premium per pet/month
- Vet clinic SaaS dashboard — recurring subscription per clinic
- Insurance partner integrations — referral fees
- Telehealth consultations — per-visit billing
- Wearable device partnerships — revenue share
- Country expansion — beyond the Philippines
THE TOOL
The framework underneath it
QuadKit is our open-source Python framework for AI backends — typed protocols, an explicit boot graph, an ASGI web layer, a CLI and a testing harness, shaped so a coding agent can operate it end to end.
GET STARTED
Two doors into QuadKit
Two ways in: start from the terminal with your coding agent, or build the graph on the Playground canvas and export the project it generates — both end in a real, standalone QuadKit project.
FROM THE TERMINAL
01 Build with your coding agent
A Python backend an LLM can actually operate: 590+ protocols, 240+ error codes, an explicit boot graph, and docs that ship in the format agents already fetch. Not autocomplete — build.
FROM THE CANVAS
02 Drag, wire, hand off
The canvas is where a project starts, not where it ends. Wire the graph, generate a complete QuadKit project, and let your coding agent take it from there — with the structure already decided.
No signup required. Hand the export to your agent with the bootstrap prompt below — same framework, same operating rules, one head start.
I'm building a Python backend with the QuadKit framework. First read https://oridecon.dev/agents.md and follow it as your operating manual, then help me scaffold the project and build my feature.
QuadKit never duplicates its rules. One sentence, one URL: the operating manual is oridecon.dev/agents.md.
THE STUDIO
Built by one person
Oridecon Labs is one person and a set of coding agents. What makes the output hold up is that the decisions stay human — and that the building kit and the products are built side by side.
- Solo · by choice
QuadKit was built alone — no team, no funding, no deadline. Everything on this page exists because one person held the judgment and kept shipping.
- Whole distance
Python services, contracts and containers at one end; React consoles and shipped surfaces at the other — one shape at both ends. The whole distance, not half of it.
- Owned · end to end
We build it and can handle the deployment — multi-tenant, RBAC-gated, audited, metered. Real deploys, real failure modes handled.
- Written down
The framework and its contracts are public; the rest stays in-house, powering the products. Writing the reasoning down is the habit this page is making a claim about.
The longer accounting — why the framework looks the way it does, and what solo-building it actually involved — is in Why I built QuadKit.
THE NOTES
Writing from the studio
How the framework is built, how agents read it, and what is stable in 0.0.x — in the open.
THE ASK
If you have something that needs building
We work best with teams at a comparable stage. And if one of our four MVPs is what you actually need, it is already running — take it further instead.
One person, no hand-off. Write to [email protected] and it reaches the person doing the work.