Applied AI engineer — Dallas, Texas

Production AI in workflows that cannot be wrong.

Applied AI engineer shipping production LLM systems into regulated enterprise workflows — automations live across Accounting, Tax, Legal, HR, Digital Marketing, and Asset Management at a 900-employee REIT.

See selected work Download the resume
Jay Singhvi

Senior AI Analyst, Welltower Inc.
MS Computer Science, Seattle University.

~1 FTE
recovered annually by automating payroll tie-out
~$100K
use tax penalty exposure resolved
#1
most-used internal GPT company-wide, six months running
6
departments running systems I shipped

Governing design principle

Arithmetic and reconciliation stay in deterministic code. The model is confined to extraction and classification.

Every system below is built so that an incorrect model inference cannot post to a ledger, file a return, or answer a policy question it does not have grounds for.

Selected work — Welltower Inc.

Eight production systems, six departments.

Each one owned end-to-end, from department requirements through UAT, deployment, logging, and failure monitoring.

Webb — enterprise HR assistant

Human Resources · ChatGPT custom GPT · OpenAI Compliance API · Snowflake

The company-wide HR assistant, migrated off a legacy Slack bot and hardened against adversarial use.

85%
fewer HR support tickets
Read the Project Details Close
  • Cut HR support tickets 85% by migrating the company-wide HR assistant from a legacy Slack bot to a ChatGPT custom GPT — 700 users (78% of a 900-employee organization), 200 monthly active users, 100+ daily queries, and the #1 most-used custom GPT company-wide for six months.
  • Held the assistant secure through emerging adversarial use across successive prompt revisions — injection red-flag patterns, identity lock, silent refusal of system-prompt disclosure, web search narrowed to checklist-gated cases, and the connector scoped to user identification.
  • Engineered out the silent-wrong-answer failure mode — a location gate withholds answers until the employee's region is known, routing only accurate information, and barring extrapolation from partial documents.
  • Shipped every revision defect-free by gating release on a regression suite and a 10-day Human Capital hallucination audit — cases pinning both correct and specific wrong responses across policy, regional routing, refusals, and injection attempts.
  • Built the AI governance layer on the OpenAI Compliance API, landing conversation data in Snowflake with a review frontend and human-in-the-loop flagging, now HR's standing weekly review.

Journal entry tie-out automation

Accounting · n8n production workflows

Payroll cash tie-out, journal entries, and suspense clearing — designed so the automation is structurally incapable of mutating the general ledger.

~1 FTE
recovered annually
Read the Project Details Close
  • Recovered ~1 FTE annually by cutting payroll tie-out from 8 people at 6 hours weekly to a single exception reviewer at 4–5 hours — implementing payroll cash tie-out, payroll journal entries, and suspense item clearing as n8n production workflows, after documenting 25–30 undocumented processes across 5–6 departments as machine-readable specifications.
  • Designed reconciliation workflows structurally incapable of mutating the general ledger — phase one proposes changes (missing chart-of-accounts entries, mapping corrections) for human approval rather than applying them, making an incorrect automated inference impossible to post.

Use tax compliance automation

Tax · n8n · Thomson Reuters tax code data

Use tax determination for state, city, and county codes across 3,000+ invoices monthly and 1,500+ properties.

~$100K
penalty exposure resolved
Read the Project Details Close
  • Resolved approximately $100K in use tax penalty exposure and established the standing control preventing recurrence across all operating jurisdictions — automating use tax determination for state, city, and county codes for 3,000+ invoices monthly across 1,500+ properties.
  • Eliminated a hallucination risk unacceptable in a compliance calculation — rebuilt an Azure live-search-per-line-item design as an n8n workflow grounded in Thomson Reuters tax code data: a reference-data lookup handles known commodities, unmatched lines escalate to a higher-capability model returning a cited proposal, and line-total mismatches route to human review.

Insurance requirements extraction and compliance benchmarking

Legal · Amazon Bedrock · document intelligence

Replaces clause-by-clause insurance review in acquisition diligence, then benchmarks every dollar limit against historical deal statistics.

Auditable
every limit traced to its source clause
Read the Project Details Close
  • Replaced manual clause-by-clause insurance review in acquisition diligence for the Legal team — document-intelligence workflows read purchase and sale agreements and extract every requirement with its dollar limits, then a deterministic benchmark engine flags each limit green, yellow, or red against historical deal statistics.
  • Made every extracted limit auditable to its source clause — a three-stage LLM chain on Amazon Bedrock handling deal context, clause extraction with verbatim quotes and page citations, then normalization of free-text limits onto standard coverage types, each stage grounded with anti-hallucination rules.

Scope creation tool

Capital projects · guided generation · Word formatting preservation

Turns a project manager's answers into complete construction scope-of-work documents. Used by every project manager on the team.

1,022
item scope corpus across 41 trade categories
Read the Project Details Close
  • Replaced manual per-project drafting for the capital projects team with a guided generation tool that turns a project manager's answers into complete construction scope-of-work documents — now used by every project manager on the team.
  • Kept every generated spec traceable to precedent and correct for its site — grounded generation in a 1,022-item scope corpus and 16 templates across 41 trade categories, conditioned per trade on a 2,384-property master dataset, then an automated climate and wind pass rewrites specs for the building's environment.
  • Preserved Word numbering, styles, and run-level formatting through an LLM rewrite — a similarity-based merge re-aligns generated text onto the original paragraph objects and keeps the original wherever confidence is low.

Performance view chat — natural language to SQL

Investment · Snowflake connector · Claude plugin

Removed the Investment team's dependency on the data team for ad-hoc analysis. The Streamlit UI was later retired for a Claude plugin.

1,500+
properties, queried in natural language
Read the Project Details Close
  • Built a natural-language-to-SQL system for the Investment team across 1,500+ properties, removing their dependency on the data team for ad-hoc analysis — later retiring the Streamlit UI for a Claude plugin pairing a Snowflake connector with skills encoding the team's output formats and report structures, so analysis runs directly in chat.

Property site map — data-driven rent setting

Asset Management · custom GPT · Excel deliverable

Generates review-ready Excel site maps for unit-level rent setting. Now the standing pricing framework for asset management.

A/B tested
beat the incumbent on occupancy, income, and preference
Read the Project Details Close
  • Built a custom GPT generating review-ready Excel site maps for unit-level rent setting — a controlled A/B test across two structurally identical halves of one property showed generated pricing beat the incumbent on occupancy, rental income, and resident preference; asset management adopted it as the standing pricing framework.

TalkFurther call center analytics

Operations · in-house call analysis · Power BI

Replaced a third-party batch transcript vendor with daily internal review of 3,000+ calls across 50+ properties.

15%
additional leads recovered from dropped conversations
Read the Project Details Close
  • Recovered 15% additional leads from dropped conversations and drove a 3% revenue increase across 50+ properties — replaced a third-party batch transcript vendor with an in-house AI analysis system moving 3,000+ calls to daily internal review, now property managers' standing KPI via Power BI dashboards linking call topics to occupancy and conversion.

Research projects — github.com/jay-singhvi

Retrieval, evaluation, and the cost of being wrong.

Resonate AI chatbot

View on GitHub

Python · LangChain · Transformers · FAISS · Pinecone · QLoRA · AWS Transcribe

90%
BERT similarity, 89% precision and recall
Read the Project Details Close
  • Reached 90% BERT similarity and 89% precision/recall with a LangChain RAG system — semantic graph clustering and overlapping chunking that preserved context across boundaries, over a FAISS and Pinecone embedding layer holding 85% cosine similarity.
  • Reduced response latency by 35% with a distributed inference and caching layer, and cut model cost via semantic routing across OpenAI, Anthropic, and Google models — evaluated on hallucination rate and response accuracy, with Llama 2 (7B) QLoRA fine-tuning cutting compute requirements by 50%.

Earlier research and applied ML work — asthma onset prediction, agentic synthetic data generation, agricultural computer vision, and marketing optimization — lives in the public repositories.

Browse all repositories

Experience

Nine years across data, research, and applied AI.

Aug 2025 – Present

Welltower Inc.
Dallas, Texas

Senior AI Analyst — applied AI engineering

  • Shipped production AI systems adopted across six departments — ChatGPT custom GPTs, Streamlit applications, and n8n workflow automation — owning each end-to-end from department requirements through UAT, deployment, logging, and failure monitoring.
  • Architected the shared automation platform every department workflow runs on — a queue orchestrator with a reusable sub-workflow library, atomic retry, and centralized error recovery, so adding a new department automation requires no orchestrator change.

Sep 2022 – Aug 2025

Seattle University
Seattle, Washington

Data Scientist — research assistant

  • Predicted asthma onset at 88% accuracy — 20% better than traditional methods — with transfer-learning and PySpark ensemble models on sparse medical data, improving forecasting accuracy by 12% over standard neural networks, on HIPAA-compliant Docker and AWS pipelines.

Apr 2019 – Jul 2022

Yardi Systems
Dubai, United Arab Emirates

Data Engineer

  • Cut implementation time and deployment cost 75% for real estate BI modules with SSIS and SQL Server ETL pipelines, and accelerated customer onboarding 60% with multi-format ingestion (batch, real-time, historical) partitioned for 40% faster processing.
  • Reduced data inconsistencies by 90% with real-time streaming BI dashboards and saved 20 hours weekly by automating reporting in T-SQL stored procedures, triggers, and tuned indexes.

Nov 2016 – Mar 2019

Yardi Systems
Pune, India

Data Engineer

  • Secured $3M+ in revenue retention and lifted system usage 50% with SSIS ETL systems that streamlined lease approval workflows.
  • Powered real-time dashboards across 100+ KPIs with a dimensionally modeled data warehouse and data marts — 15% faster processing via incremental loading — and set the governance standards, automated testing, and query tuning behind them.

Technical skills

The stack behind it.

AI and machine learning

Python · PyTorch · RAG · LLM evaluation · embeddings · vector databases (FAISS, Pinecone) · prompt engineering · guardrail and anti-hallucination design · OpenAI (GPT, custom GPTs, Compliance API) · Anthropic Claude (plugins, skills, MCP) · Amazon Bedrock · LangChain · Hugging Face · Scikit-learn

Data and analytics

Snowflake (Cortex Analyst) · SQL · T-SQL · PostgreSQL · PySpark · Spark SQL · Pandas · NumPy · ETL pipeline development · SSIS · dimensional modeling · data warehousing · incremental loading · query optimization

Cloud and infrastructure

AWS (S3, SageMaker, Transcribe, Bedrock) · Docker · Git · GitHub Actions CI/CD · RESTful APIs

Applications and reporting

Streamlit · n8n · Power BI · Tableau · KPI dashboards · Excel

Education

Education

MS, Computer Science — data science

Seattle University, Seattle, Washington · Sept 2022 – Jun 2024

MS, Computer Applications

Symbiosis International University, India · July 2015 – Apr 2018

BS, Information Technology

University of Mumbai, Mumbai, India · Jun 2011 – Jan 2015

Research papers — /publications

Publications and certifications

DAWAK 2024

Incremental SMOTE with Control Coefficient for Classifiers in Data Starved Medical Applications

SAC 2025

A Retrieval-Augmented Framework for Meeting Insight Extraction

Peer review — IEEE JBHI 2025

Hybrid Deep Learning using Transfer Learning as Feature Extractor in Environmental Health Risk Prediction

CITI Program

Responsible Conduct of Research – Engineers · Human Subjects Research for IRB (Faculty, Staff, and Student)

Other certificates on LinkedIn

Get in touch.

Open to applied AI engineer, forward deployed engineer, and process-to-AI automation roles.

206-660-3609
LinkedIn GitHub Google Scholar Resume

This opens a prefilled draft in your own mail client — nothing is sent from the page.