VP / Director of Engineering · Agentic AI · MCP · MLOps · CPQ · Quote-to-Cash · Hands-On Technical Leader

Belmont, CA · He/Him

I build AI/ML platforms that generate revenue — not just predictions. Hands-on engineering leader who architects and codes production systems personally while setting technical direction for a 40+ person global org.

17+ years at Oracle shipping two flagship AI platforms at enterprise scale:

• Agentic AI on CPQ — Renewal Agent + Quote Generation Agent · 3,000+ active sales users · 600+ enterprise clients · 50+ countries · 30% renewal cycle compression · 28% quote processing efficiency improvement
• Unity CDP AI/ML — 6 production models (Next Best Action, Churn Propensity, CLV, RFM Segmentation, Multi-Touch Attribution, MMM) · 9,000+ customers at day-one GA · no phased rollout

The hard part isn't the model — it's the integration layer. I solved the cross-vendor problem in production: unified live context from Salesforce, MS Dynamics 365, and Oracle clouds so agents make decisions on real data, not cached snapshots.

Production-grade, cloud-agnostic Agentic AI platform for banking decisions — complete reference architecture for governed agentic AI systems. Ten platform capabilities across eight services, composable by any product team through a stable SDK.

"The engineering problems that make agentic AI fail in production are not model problems. They are infrastructure problems — stale context, ungoverned execution, absent memory, unvalidated models, and no feedback path from outcome back to decision. This platform solves each of those problems as a named, typed, independently testable layer with a clean contract to its neighbors."

Three architectural deficits this platform addresses:

• Stale batch context — nightly risk scores reflect yesterday's account state. An agent acting on an 18-hour-old risk score makes a decision that is technically correct but wrong in the world.
• Ungoverned agent execution — an agent with no compliance gate between its reasoning and its action can violate CFPB, ECOA, or UDAAP. In a regulated environment, that is not a product risk — it is a legal one.
• No closed-loop governance — without outcome capture, memory, and evaluation history, the platform learns nothing across sessions. It restarts blind every time.

Six architectural principles — applied without exception:

1. Typed contracts at every boundary — Pydantic v2 schemas throughout, no dicts, no untyped kwargs
2. Protocol-based dependency injection — every external dependency behind a Protocol interface, independently testable
3. Graceful degradation over hard failure — one failing source marks sources_degraded, pipeline continues
4. Governance as a runtime capability — Layer 4 runs before Layer 6 executes, not as a post-hoc audit
5. Immutable, replayable audit trail — one trace_id reconstructs every decision for regulatory replay
6. Closed feedback loop — outcome events write CustomerMemory records, retrieved at the next session

Reference Architecture — Six Governed Layers:

One trace_id reconstructs every decision end-to-end for regulatory replay. 4-gate offline evaluation pipeline: benchmark · fairness · Adverse Impact Ratio · LLM-judge.

10 UI pages: Pipeline Runner · Architecture View (animated SSE) · Audit Trail · Experiments · Drift Monitor · Guardrails · Model Registry · Evaluation · Settings · About

LLM modes — runtime switching, no restart required:

8 local services:

Valkey:6379  PostgreSQL:5432  Qdrant:6333  Jaeger:16686
Prometheus:9090  Grafana:3000  MLflow:5001  Ollama:11434

git clone https://github.com/saralabiswal/agentic-banking-llmops
cd agentic-banking-llmops && make install && make docker-up
cp .env.example .env && make seed && make dev
# All 8 services + API + UI with hot-reload — no API key required

MCP-powered integration layer for vendor-agnostic quote-to-cash agentic decisions — the cross-vendor architecture pattern running in production across 600+ enterprise clients.

The core proof: vendor selection is configuration, not agent code. Switching CRM from Salesforce to Microsoft Dynamics 365, or Order Management from Oracle FOM to SAP S/4HANA, changes the adapter path and source attribution — the decision agent and canonical schema are untouched.

16 adapters. 5 commercial-system slots. One canonical schema. Seven demo scenarios.

git clone https://github.com/saralabiswal/agentic-mcp-quote-to-cash
cd agentic-mcp-quote-to-cash && make install && make seed
make dev-api          # FastAPI → http://localhost:8000
cd ui && npm install && npm run dev -- --port 3001
# No API key required — runs end-to-end in demo mode

Evaluation harness for governed LLM agents — because production AI without evals is just a demo you shipped.

Most agentic AI systems stop at building the agent. This framework answers the question every production deployment eventually faces: how do you know it's still working correctly next month?

Plugs directly into agentic-banking-llmops as its evaluation layer — same trace_id, same scenarios, same policy boundaries.

make install && cp .env.example .env
make demo                         # mock backend, no API key required
make dev                          # API → http://localhost:8001

Production LLMOps control plane for Quote-to-Cash agentic workflows — token cost attribution, quality scoring, latency SLOs, prompt versioning, and semantic drift detection across five providers.

Answers the production questions most enterprises cannot answer:

Five providers with real rate cards:

make install && make seed
ollama pull llama3.2              # default local model
make dev-api                      # API → http://localhost:9100
make dev-ui                       # UI  → http://localhost:5173
# No API key required

AI / ML / Agentic

Languages & Frameworks

Cloud & Infrastructure

Data & ML Stack

• Post Graduate Diploma, Machine Learning — Cornell University, NY
• MBA, Technology Management — University of Phoenix, AZ
• B.S., Computer Science & Engineering — Utkal University, India

I build the platforms that make AI commercially accountable — not just technically impressive.