Production Agentic Systems, Owned End-to-End
Rupam Das (@rejisterjack)
I treat agentic AI as a production systems problem, not a demo. For 3+ years I have owned the path from distributed backends to LLM products running on top of them — including a multi-tenant marketplace with defense-in-depth isolation and an autonomous tool-calling agent used by real users.
I lead from the codebase: a 5-engineer team at EL Nova Labs, CI that refuses to ship unscoped tenant queries, and a 2,000+ line golden-eval suite that gates every Eva release at PostQode.
Foundational Engineering Disciplines
Operating Principles
Isolation is a feature.
Tenant boundaries are enforced in Prisma middleware, in SQL Server Row-Level Security, and again in CI. If a query can leak across tenants, it never reaches production.
// Defense-in-depth Prisma tenant scoping middleware
prisma.$use(async (params, next) => {
if (TENANT_SCOPED_MODELS.has(params.model)) {
const tenantId = getSessionTenantContext();
if (!tenantId) throw new SecurityBoundaryError("Unscoped tenant access");
params.args.where = { ...params.args.where, tenantId };
}
return next(params);
});Agents without evals are prototypes.
Tool-calling systems ship with golden evals, PII scrubbing, citations, and circuit breakers — or they do not ship. We evaluate tools with strict JSON schemas and merge-blocking assertion suites.
// Golden eval assertion gate for autonomous tool execution
describe("Autonomous Agent Tool Gate", () => {
it("strictly validates parameter bounds before model execution", async () => {
const payload = sanitizeAgentInput(rawInput, { scrubPII: true });
const toolCall = await evalRunner.executeScenario("cart_checkout_flow");
expect(toolCall.schemaValid).toBe(true);
expect(toolCall.piiLeakDetected).toBe(false);
});
});Latency is a product decision.
A 93% cut (3s → 200ms) and a 40% load-time reduction are architecture, not micro-optimizations. Real-time experiences demand off-thread computation, edge streaming, and proactive cache topologies.
// Low-latency chunk streaming with early speculative cache
export async function streamAudioOrchestrator(req: Request) {
const cached = await redis.get(cacheKey);
if (cached) return new Response(cached, { headers: STREAM_HEADERS });
const stream = createMultiplexedStream(aiAudioSource, {
chunkSize: 1024,
energyGateThreshold: 0.015,
});
return new Response(stream, { headers: STREAM_HEADERS });
}The pipeline is part of the product.
Trivy, SBOM, cryptographic image signing, OpenAPI/DB drift checks, and blue-green deploys are how trust is built. Software that cannot be verified cannot be trusted in enterprise production.
// CI deployment gate: Container signing and drift verification
jobs:
verify-and-sign:
steps:
- name: Scan vulnerability baseline
run: trivy image --severity HIGH,CRITICAL --exit-code 1 $IMAGE_TAG
- name: Sign container artifact
run: notation sign --key-id $NOTATION_KEY $IMAGE_TAG
- name: Verify zero database drift
run: bunx prisma migrate diff --exit-codeOpen Source & Public Deliverables
Verified Production Artifacts
Enterprise RAG Starter Kit
Production-grade self-hosted RAG platform over 10,000+ enterprise documents with hybrid vector search, Reciprocal Rank Fusion (k=60), and pgvector.
Frontend Dev Helper
Manifest V3 Chrome extension with 19 injected tools replacing eight legacy utilities. Off-thread DOM inspection and zero layout shifts.
Nexus Dual-Modality AI
Interactive voice agent running OpenAI Realtime over WebRTC + streaming portfolio RAG assistant with energy-gated duplex barge-in controls.
65+ Developer Tools Sandbox
Comprehensive browser-side utility sandbox executing cryptography, formatters, and transforms locally in V8 with zero network telemetry.
First-Party Facts & Citation Guide
Architectural FAQ & Verification
Rupam Das (known online as rejisterjack) is a Full Stack AI Engineer based in Bengaluru, India. He specializes in treating agentic AI as a production systems problem: multi-tenant isolation, tool-calling evals, and supply-chain-hardened delivery. He has owned systems at PostQode, led a 5-engineer team at EL Nova Labs, and engineered real-time telemetry at Big Wings LLC. All roles and metrics are grounded in verified first-party artifacts.
Interview Rupam or Question His Architecture Live
Have questions regarding two-layer tenant isolation, golden eval merge gates, or real-time WebRTC audio pipelines? Test his architectural rationale with Nexus via streaming chat or low-latency voice.
Ready to ship resilient, production-grade agentic systems?
Whether you need staff-level ownership for multi-tenant AI backends, golden eval CI gates, or low-latency streaming architectures, let’s build systems that scale.