Case study · agent infrastructure · verified public release
Agentic Systems Gateway provides one policy boundary for zero-cost replay, optional visitor-owned NVIDIA NIM or OpenRouter inference, system discovery, release evaluation, signed receipts, and operational telemetry.
Agentic Systems Gateway demonstrates how a public portfolio can expose real, input-sensitive system behavior without storing visitor keys, lending an owner-funded provider credential, accepting arbitrary model endpoints, or confusing release evidence with client production acceptance.
Case-study figures describe this documented engagement and are not forecasts or guarantees.
A public agent portfolio creates a difficult operating boundary. Visitors need a working path when free providers are unavailable, live model calls cannot create an open-ended owner bill, API keys must not enter logs or storage, and a passing test suite must not be presented as evidence that a client accepted a production integration.
Aixcel built a strict FastAPI and Pydantic control plane with a nine-system registry, nine evidence-qualified evaluation records, fixed NVIDIA NIM and OpenRouter provider adapters, deterministic evidence replay, request quotas, prompt-injection checks, HMAC-signed receipts, an optional redacted SQL ledger, generated OpenAPI, and a persistent light and dark workspace. LangGraph is intentionally excluded because this component is a deterministic policy boundary rather than an agent reasoning workflow. The browser sends a visitor key in one request header and clears the field after use. It cannot supply a model name, endpoint, tool, or external action. SQLAlchemy, Alembic, PostgreSQL, OpenTelemetry, Prometheus, Structlog, Docker, GitHub Actions, Postman, Playwright, and a pinned Linux Vercel builder complete the release path.
The public deployment exposes nine systems and their tested release boundaries. Replay works without a model account and changes when evidence changes. Live NVIDIA NIM and OpenRouter paths require a visitor-owned key. Shared NVIDIA access exists as a disabled code path and cannot create spend until a durable cross-instance quota store, approved key, monitoring, and an explicit cost decision are supplied. The verified release passed 85 tests at 84.95 percent measured coverage, 18 of 18 evaluation measures, 12 production requests with 40 Postman assertions, four durable redacted receipts across a PostgreSQL API restart, desktop and 390-pixel browser journeys, seven rendered Swagger operations, persistent themes, zero browser errors, and zero production 5xx responses.
Python 3.12, uv, FastAPI, Pydantic v2, httpx, SQLAlchemy, PostgreSQL 17, Alembic, REST, OpenAPI, Postman, HMAC receipts, OpenTelemetry, Prometheus, Structlog, Docker, GitHub Actions, Playwright, Vercel
A framework earns its place by owning a clear responsibility in the system, not by appearing in a technology list.
What it is: FastAPI is an async Python API framework. Pydantic validates and serializes typed contracts.
Why it is here: They reject malformed input and generate the same OpenAPI contract used by the browser, Postman, and external reviewers.
What it is: httpx is an async HTTP client for Python.
Why it is here: It provides bounded timeouts and explicit request construction for the two fixed provider adapters without embedding provider behavior in the UI.
What it is: SQLAlchemy maps Python records, Alembic versions schema changes, and PostgreSQL provides durable relational storage.
Why it is here: The private deployment path can preserve redacted, transactional receipts while the public Vercel release keeps persistence disabled.
What it is: OpenTelemetry standardizes traces, Prometheus exposes time-series metrics, and Structlog emits contextual JSON logs.
Why it is here: Together they make requests, provider behavior, limits, security events, and failures inspectable without committing to a paid monitoring vendor.
What it is: Postman is a visual and command-line API testing platform for collections, environments, and assertions.
Why it is here: The 12-scenario collection proves health, capabilities, registry, evaluation, replay, sensitivity, injection blocking, and the no-key live boundary against production.
What it is: Docker packages a reproducible service and PostgreSQL runtime. Vercel provides on-demand serverless HTTPS deployment.
Why it is here: Docker proves the durable private path while Vercel keeps the public synthetic gateway available without paid idle compute.
Each component owns a specific responsibility. Alternatives were rejected only where they added complexity or weakened the tested control boundary.
| Responsibility | Choice | Why it fits | Alternative and constraint |
|---|---|---|---|
| Public API | REST with FastAPI and Pydantic | Clear resources, typed errors, generated OpenAPI, and direct browser and Postman testing | GraphQL adds query and authorization surface; gRPC is less useful for a public browser workspace |
| Policy flow | Deterministic Python, not LangGraph | Provider allowlists, quotas, key handling, receipts, and security decisions should be direct and reproducible | LangGraph is valuable inside stateful agent workflows but would add orchestration where this gateway needs a fixed policy boundary |
| Free demonstration | Verified deterministic replay | The system remains inspectable during provider outages and produces no inference bill | An owner-funded public model key creates cost and abuse risk; a fixed canned page would fail the input-sensitivity requirement |
| Live access | Visitor-owned keys for fixed providers | The visitor controls provider quota while the gateway controls model, endpoint, token, timeout, and tool boundaries | Arbitrary OpenAI-compatible endpoints or model names would expand SSRF, cost, and unsupported capability risk |
| Durable evidence | HMAC receipts and optional redacted PostgreSQL ledger | Reviewers can verify provider and content hashes without persisting prompts, keys, or raw output | Full prompt logging creates unnecessary privacy and credential risk |
| Deployment | Vercel public release plus Docker and PostgreSQL private path | The public interface scales down between requests while the container path proves durable restarts | Kubernetes is deferred until traffic, tenancy, or a buyer requirement justifies cluster operations |
The system is credible only when its input limits, release tests, and operating signals are visible together.
The gateway does not train a model. Its controlled release corpus contains nine synthetic system records, nine evaluation records, and fixed replay scenarios for decision explanation, evidence critique, and structured summarization. Records disclose the live URL, data boundary, mutation boundary, tested scope, release proof, and replay scenario for each system. No client credentials, prompts, provider keys, patient information, private records, or raw model output enter the persistent receipt schema.
The release suite checks registry completeness, evaluation coverage, passing scores, zero external writes, disclosed data and mutation boundaries, replay availability, fixed HTTPS providers and models, disabled shared spend, redacted receipt schema, theme persistence, Postman contracts, PostgreSQL restart behavior, and the coverage floor. Unit, schema, security, provider, replay, registry, API, generation-drift, Postman, Docker, desktop, mobile, input-sensitivity, key-storage, Swagger, and runtime-log checks form the wider acceptance gate.
Every response receives a trace ID. OpenTelemetry propagates request and provider spans. Prometheus exposes more than 15 signal families across HTTP traffic, latency, provider calls, quotas, injection blocks, live and replay modes, errors, signed receipts, persistence, and registry access. Structlog emits contextual JSON without prompt or key content. Suggested alerts cover readiness, 5xx rate, p95 latency, provider error ratio, quota denials, security blocks, receipt failures, and any unexpected shared-provider activity.
Names and sensitive details are withheld. Metrics retain their stated meaning and evidence label. A scope count is not converted into an outcome, and no engagement result is presented as a universal benchmark.
Compare adjacent systems, inspect evidence, or see how Aixcel delivers the work.
In 25 focused minutes, we will map where work or revenue is getting stuck, test whether AI is the right intervention, and identify the highest leverage first step.
Book a free systems audit