Eleventh turns fragile retrieval, agent, and AI-backend prototypes into systems that survive real users, real data, and real failures.
Each row is one engineering capability. The diagram on the right is how the system runs in production. Every row links to its source repo and (when applicable) the live deploy. Read the source before the first call.
hybrid retrieval + rerank
Your retrieval works in demos but fails on real user queries. Precision drops, irrelevant results surface, and answer quality degrades under messy, unstructured input.
Hybrid retrieval (pgvector + BM25) with reranking, instrumented end-to-end. Built for the gap between a demo that retrieves and a system that survives the next 10,000 queries.
state graph + approval gate
Your agent works on the demo path but breaks when users deviate. State gets lost between steps, tool calls fail silently, and there is no way to audit what happened.
Multi-step agent workflows on LangGraph with persistent state, tool orchestration, approval gates, retry logic, and deterministic evaluation. Designed for reliability under real interaction patterns.
request latency waterfall
Your AI feature needs a real backend, not a notebook wrapped in an endpoint. You need async APIs, structured data access, migration discipline, and deployment automation from day one.
Asynchronous FastAPI backends purpose-built for LLM-powered apps. pgvector, Alembic migrations, health checks, and operational readiness on day one.
regression detection · 12 runs
You are scaling prompts before you have evaluation discipline. Quality is checked by eyeballing outputs. Cost and latency are unmeasured. When something breaks in production, there is no way to tell what changed.
Deterministic evaluation frameworks, structured output validation, cost and latency observability, regression detection. The layer most teams skip between prototype and production.
Every frame below is the real system, captured from its live deploy. One runs in production, three run public benchmarks, two are honest showcases. All six are open source and running right now. Click any frame and use the real thing.
flagship · Production
Multi-tenant RAG platform
Multi-tenant, multi-cloud RAG platform with SSO, SCIM, RBAC/ABAC, envelope encryption, multi-region failover, and SOC 2 automation. The shipped reference every other repo points to.
// supporting fleet · every frame is the system, live
// captured from the live deploys · June 2026 · see the full fleet, by stage
Three engagement profiles. Each starts with a measurable problem, ends with deployed code in your repo, and gets handed off with the eval harness, observability, and the docs we used to ship it.
FOR ·01
You demoed the AI feature. Now the team is asking for retrieval that doesn’t fall over, an eval harness with regressions caught, and an API your iOS engineer can actually call.
FOR ·02
Your team builds the front end and the brand. We ship the AI backend underneath: RAG, agents, eval. Then we disappear before launch with the source in your client’s repo.
FOR ·03
You’re considering an AI investment and want a working reference architecture, an eval against your real data, and a number on what production reliability actually costs.
Every engagement runs through the same sequence. Discovery sets the spec and the evaluation bars; the next four phases build and ship against them. Adjustable by week, not by order.
We sit with the brief, write the system spec, define the evaluation bars, and surface every risk before any code ships. The spec we write here is the contract every phase below is measured against.
// deliverables4 files · spec-only
Every system ships a public stats endpoint. The console below probes all six from your browser every 30 seconds and prints what comes back, including the measured round-trip of each request. When a system is quiet, it says so. Nothing here is simulated.
Multi-cloud RAG · production workload
GET /api/stats → …
// request log · fetched live from this browsernewest last
--:--:--Zawaiting first responses…
The difference is not the model. It’s the system around it. Below are the four laws we won’t ship without. Each is paired with the anti-pattern that gives the principle its enforcement edge.
Law 0101 / 04
Replacing subjective “looks good” testing with measurable scoring and regression checks. Every prompt change ships with a delta against a frozen baseline.
Law 0202 / 04
Preventing uncontrolled token usage through architecture, not afterthought optimization. Cost ceilings, caching, and routing live at the request layer.
Law 0303 / 04
Designing systems that respond in real time, not seconds too late. p95 is a feature; warm-paths and streaming are architectural, not optimization.
Law 0404 / 04
Building systems that continue working under load, failure, and scale. Health checks, retries, idempotency, and circuit breakers ship on day one.
//position
AI infrastructure, not AI theater.
// from the engineering charter
Tell us about the system you need shipped. We respond in 24 hours with a specific technical plan, a measurable bar, and an honest estimate of weeks and cost.
Brief us in two paragraphs. If we’re not the right fit, we’ll tell you fast and route you to a team that is.
// prefer email? hello@eleventh.dev
