Amey Parmarthi

Forward Deployed Engineer · AI Deployment Leader · Salesforce

I ship enterprise AI agents
to production — and build
the infrastructure underneath.

I lead production Agentforce deployments for Fortune 500 customers — owning agent architecture, evaluation, and the latency fixes that get them live. Then I go build the agentic systems myself: an MCP server, a multimodal financial RAG engine, production ML. Fifteen years carrying AI from the boardroom to the codebase.

What a Forward Deployed Engineer does

strategy + shipped code
01

Ship agents to production

I lead enterprise Agentforce deployments end to end — agent architecture, prompt design, evaluation frameworks, and production reliability across Fortune 500 accounts. When a deployment hit critical latency, I traced three compounding causes and drove the prompt-migration and payload-filtering fixes.

02

Build the infrastructure

I don't hand off and hope. I build the agentic systems and ML underneath — an MCP server connecting Claude to MLflow, a multimodal financial RAG pipeline, production models behind APIs — with evaluation harnesses I treat as first-class, not an afterthought.

03

Translate both directions

Fifteen years translating C-suite objectives into technical roadmaps and back. I can define KPIs and manage trade-offs with executives, then go debug a serving pipeline — the dual fluency that turns AI pilots into production revenue.

Experience

15 years · Deloitte → Salesforce

Forward Deployed Engineer / AI Deployment Leader

Salesforce — FDE Organization
2026 — Present
  • Lead technical delivery of enterprise Agentforce deployments — agent architecture, prompt design, evaluation frameworks, and production reliability across Fortune 500 accounts.
  • Diagnosed and resolved critical latency in a production deployment — isolating three compounding causes (token-streaming blocked by JSON-only architecture, oversized LLM payloads, slow baseline API) and driving the fixes.
  • Translate ambiguous enterprise workflows into multi-step agent flows across Agentforce, Data Cloud, Flows, and external APIs; operate as the field-to-product feedback loop into the platform roadmap.

Sr. Manager, Solution Architect — AI / CRM / Data Cloud

Salesforce
2015 — 2025
  • Delivered Agentforce pilots for enterprise customers — prompt patterns, safety guardrails, and agent success criteria for high-volume service environments.
  • Architected Data Cloud–backed Customer 360 solutions integrating 20+ systems for real-time identity resolution and personalization at scale.
  • Translated C-suite objectives into technical roadmaps across multi-year programs spanning financial services, healthcare, technology, and consumer electronics.

Senior Consultant — Salesforce Program Lead

Deloitte Consulting
2011 — 2015
  • Owned end-to-end architecture, data modeling, integration design, and go-live for enterprise Salesforce programs at Fortune 500 clients.
  • Partnered with executive stakeholders to define roadmaps, manage trade-offs, and connect technical delivery to measurable business outcomes.
Education
  • M.S. Computer Science & EngineeringUniversity at Buffalo (SUNY)
  • AI/ML Engineer Career TrackSpringboard × Univ. of Arizona · 2026
  • Product & Agile ManagementUC Berkeley Executive Education
Certifications
Agentforce SpecialistData Cloud ConsultantApplication Architect
Status

U.S. Permanent ResidentNo sponsorship required

Engineering Projects

built, not described
01
−55%
LLM cost via intent routing · in progress

FinSight — Financial Evidence Engine

in progress

A multimodal RAG system answering investor-grade questions across SEC filings, earnings transcripts, XBRL fundamentals, and OHLCV market data. Two-stage hybrid retrieval (BM25 + voyage-finance-2 dense, fused via RRF in Qdrant) with Cohere reranking. LangGraph orchestration routes by intent — Sonnet for synthesis, Haiku for classification — cutting cost ~55%. Responsible-AI guardrails throughout: PII detection, jailbreak prevention, RAGAS faithfulness scoring, citation verification.

LangGraphClaudeQdrantCohereRAGASFastAPI
source
02
MCP
6 tools · stdio + SSE · Docker

MLflow MCP Server

open source

A Model Context Protocol server giving Claude direct read/write access to MLflow experiment tracking — six tools spanning run search, logging, and tagging. Dual transport (stdio for Claude Desktop, SSE for hosted), Docker-ready, stateless by design. The connective tissue AI-native teams are racing to build.

Python 3.11FastMCPMLflowAnthropic
source
03
0.966
AUC · +$17.6K vs −$6.8K naive policy

KKBox Churn Prediction Platform

live

Churn prediction on 30M+ user interaction records — gradient-boosting ensemble (XGBoost/LightGBM) at 0.966 AUC with SHAP explainability. The real work is the decision policy: I showed the naive 'target the top 10K riskiest' approach loses money (−$6.8K), then tuned a cost-sensitive threshold that turns the same model ROI-positive (+$17.6K/campaign). Shipped with a FastAPI endpoint, MLflow registry, and a live Streamlit ROI simulator.

LightGBMXGBoostMLflowFastAPISHAPStreamlit
live demo