Girijesh Singh

The Moat

The 6-Year Thread

Most engineers change companies and domains. I stayed with the same core problem for six years and solved it with progressively more powerful technology. That's rare — and hard to replicate.

2018

Pre-LLM Era

NL-to-SQL

Before LLMs existed — fine-tuned Rat-SQL to translate insurance adjuster questions into complex SQL, automatically generating interactive dashboards from results.

Fine-tuned Rat-SQL to handle complex JOINs & WHERE clauses specific to insurance-domain schemas

End-to-end pipeline: Natural Language → SQL → Database → auto-generated interactive dashboard

Team of 10 — owned full design, development, and maintenance lifecycle

Demoed to external insurance clients; adopted internally at scale

PythonRat-SQLNLPPostgreSQL

Click to expand ↕

2023

LLM Agent Era

LLM Agents → Custom Architecture

Built with LangChain, diagnosed its limits in 30 days, redesigned the entire architecture solo. Then cut latency from 2.5 minutes to 40 seconds — in three weeks.

LangChain broke on multi-document insurance policies — no pinpoint citations, cross-document context collapse

Proposed scrapping LangChain entirely. Delivered replacement architecture solo in 3 months (one colleague pressure-tested)

Latency crisis resolved: 2.5 min → 40 sec via single-trip retrieval redesign + LLM calls reduced 5→2

This architecture became the production foundation for everything that followed

LangChainGPT-4Custom RAGFastAPI

Click to expand ↕

2024

Enterprise Platform Era

Enterprise RAG Platform

Custom hierarchical retrieval. No framework dependencies. 4 global regions. 2,000+ active users. New carriers onboarded through configuration alone — zero code.

Sections → subsections → bullets → node numbers → network retrieval (proprietary hierarchy)

Real-time citation validation — hallucinated sources eliminated character-by-character before reaching users

Graph-based execution engine: new insurance carriers onboarded via config, zero code changes required

10× latency reduction, 20× throughput gain, $10M+ estimated annual savings in adjuster time

Azure OpenAIGPT-4oMicroservicesAdobe PDF API

Click to collapse ↕

By the Numbers

Real Impact at Scale

Production AI in a regulated, high-stakes industry — every number below is earned, not estimated from a demo.

—

Active Users

Insurance adjusters · US · UK · AU · EU

—

Daily AI Queries

Multi-carrier, multi-tenant platform

—

Latency Reduction

Hours of manual work → under 30 seconds

—

Est. Annual Savings

In insurance adjuster time costs

—

Global Regions

US · UK · Australia · EU deployment

—

Time to Promotion

Senior → Lead Data Scientist

Featured Work

What I've Built

Production systems, published research, and shipped products — not demos. Everything here is real and has been used at scale.

🏗️

PRODUCTION · NDA

Enterprise Insurance AI Platform

Lead Data Scientist · 2024–Present · Confidential Employer

End-to-end enterprise RAG platform that reduces insurance policy review from 30–45 minutes to under 30 seconds. Built on a custom hierarchical retrieval architecture — no LangChain, no LlamaIndex — designed from scratch to handle multi-document, cross-referencing insurance policies (up to 700 pages) at enterprise scale across 4 global regions. Covers the full adjuster workflow: multi-engine document ingestion, coverage Q&A with pinpoint citation validation (hallucinations eliminated before reaching users), coverage verdict generation, and automated letter population. New insurance carriers onboard through configuration alone — zero code changes.

2,000+

Active Users

30K+

Daily Queries

10×

Latency Gain

20×

Throughput

$10M+

Annual Savings

Regions

Custom RAG Hierarchical Retrieval Graph Execution Engine GPT-4o Azure OpenAI Microservices Adobe PDF Services Tesseract OCR FastAPI PostgreSQL MongoDB Azure Service Bus

⚡

PRODUCTION

FNOL Claim Classification System

Sole ML Engineer · 2018–2021 · Confidential Employer

First Notice of Loss (FNOL) classification agent for a major US insurance carrier. Sole ML engineer — owned model development, client interaction, and deployment. Automated routing of 200–300 incoming insurance claims per day to the correct claim type and adjuster team. CTO called it the best thing any dedicated team had built.

30%

Labor Reduction

45%

Cost Reduction

35%

Faster Settlement

XGBoost Scikit-learn Python Insurance ML Interactive Dashboard

🔤

INTERNAL PRODUCT

PriML — Natural Language to SQL

Team of 10 · 2018–2021 · Pre-LLM Era

Before LLMs existed, built an NL-to-SQL product for insurance domain data. Fine-tuned Rat-SQL to handle the complex multi-table, multi-JOIN queries specific to insurance schemas. NL queries flowed into SQL, hit the database, and auto-generated interactive dashboards — demoed to external clients.

Rat-SQL Fine-tuning NLP Python PostgreSQL Auto-Dashboard

📄

PUBLISHED · PST 2023

Privacy-Preserving ML for Brain-Computer Interfaces

Research · Lakehead University · Copenhagen, Denmark · 2023

Published at the 20th IEEE International Conference on Privacy, Security and Trust (PST 2023) in Copenhagen. Proposes a framework for privacy-preserving machine learning applied to BCI neural data — preventing re-identification of sensitive brain activity signals while preserving model utility.

Differential Privacy BCIs PyTorch Published Research IEEE PST 2023

Technical Arsenal

What I Work With

Tools chosen for production reliability, not resume padding. Highlighted chips indicate expert-level depth.

🤖 LLMs & Generative AI

Custom RAG Architecture GPT-4o / Azure OpenAI Hierarchical Retrieval Prompt Engineering Embedding Models Citation Validation Multi-Document QA LangChain LlamaIndex Pydantic Output Schemas

📊 ML & Data Science

Python SQL XGBoost Scikit-learn PyTorch PySpark Feature Engineering Statistical Modeling Anomaly Detection NLP / Fine-tuning

☁️ Infrastructure & Engineering

Azure FastAPI Docker PostgreSQL MongoDB Redis Azure Service Bus Azure Blob Storage Microservices REST APIs

🛠️ Domain & Leadership

Insurance Domain (6 yrs) Team Leadership (10–14) Adobe PDF Services API Tesseract OCR GPT-4o Vision API DAG Orchestration Observability Stakeholder Management Git

Experience

Career Timeline

7+ years of progressively deeper AI work, from solo ML engineer to leading a cross-functional team of 14.

Jan 2024 — Present

Lead Data Scientist

Primus Software Corporation

📍 Waterloo, ON, Canada

Led a cross-functional team of 10–14 (frontend, backend, QA, 7 data scientists). Scaled enterprise AI platform to 2,000+ active users across US, UK, AU, and EU. Resolved two production crises, built the graph-based execution engine for zero-code carrier onboarding, and extended the platform into coverage verdicts and automated letter generation.

Promoted from Senior → Lead in 12 months

Jan 2023 — Jan 2024

Senior Data Scientist

Primus Software Corporation

📍 Waterloo, ON, Canada

Diagnosed LangChain's fundamental limits on multi-document insurance policies, proposed scrapping the framework, and designed the hierarchical RAG architecture from scratch — solo — in 3 months. Resolved a latency crisis (2.5 min → 40 sec) through single-trip retrieval redesign and cutting LLM calls from 5 to 2.

Sep 2021 — Apr 2023

M.Sc. Computer Science + Graduate Research Assistant

Lakehead University

📍 Thunder Bay, ON, Canada

Project-based Masters, supervised by Dr. Garima Bajwa. Published privacy-preserving ML research for Brain-Computer Interfaces at PST 2023 International Conference (Copenhagen). Continued part-time AI development at Primus Software concurrently — retained due to irreplaceability.

Published at PST 2023 · Copenhagen, Denmark

May — Aug 2022

Data Science Intern

Ciena

📍 Ottawa, ON, Canada

PySpark-based data pipelines, divisive clustering, manufacturing batch anomaly detection. Recognized early that the role skewed toward data engineering — confirmed commitment to applied data science path.

Jun 2018 — Dec 2022

ML Engineer → Senior ML Engineer

Primus Software Development

📍 Noida, India (+ Remote, Canada from 2021)

Built the FNOL classification system for Crawford & Company as sole ML engineer with direct client ownership. Built PriML (NL-to-SQL product) as technical lead on a team of 10. Insurance domain expertise begins here. Retained remotely by Primus after moving to Canada for Masters — a clear signal of value.

CTO recognition + $1,000 performance bonus

Research

Publications

📚 Peer-Reviewed · International Conference

Privacy-Preserving Machine Learning Framework for Brain-Computer Interfaces

20th IEEE International Conference on Privacy, Security and Trust (PST 2023) · Copenhagen, Denmark

Addresses critical re-identification risks in neural data processing. Proposes a framework that applies privacy-preserving techniques to BCI machine learning pipelines — preventing re-identification of sensitive brain activity data while maintaining model utility across tasks. Supervised by Dr. Garima Bajwa, Department of Computer Science, Lakehead University, Thunder Bay, ON.

Beyond the Code

Leadership & Communication

Data scientists who can communicate build better systems. The evidence is below.

🎤

Toastmasters International

Competitive public speaking training that directly informs how I present technical findings to non-technical stakeholders — executives, clients, and insurance carriers.

7× Best Impromptu Speaker

4× Best Evaluator

3× Best Prepared Speech

👥

Cross-Functional Team Lead

Ran day-to-day technical and delivery decisions for a team of 10–14. Direct stakeholder requirement gathering, refinement, and brainstorming. Second-most senior person on the team — manager above, Girijesh ran everything technical below.

10–14 person cross-functional team

Multi-region, multi-carrier delivery

Client-facing requirement ownership