PORTFOLIO — 2026 · KITCHENER, ONTARIO, CA
Lead Data Scientist & Applied AI Architect — building production AI where hallucinations carry legal consequences.
Three things that separate a production AI architect from someone who can run a notebook.
I don't just prompt-engineer; I build deterministic, multi-agent microservices from scratch. When off-the-shelf tools like LangChain fail at scale, I design custom orchestration layers (PydanticAI, FastMCP) that actually work in production.
I bridge the gap between Data Science and Data Engineering. I optimize vector indices (Azure Cosmos DB IVF) to slash query latencies from minutes to seconds, circumventing hundreds of thousands in cloud scale-out costs.
My architectures don't just live in notebooks. I've directed teams of 10–14 engineers to deploy systems that drive $10M+ in operational efficiency and process tens of thousands of queries daily across 4 global regions.
TECHNICAL PHILOSOPHY
Moving a GenAI prototype to a regulated production environment exposes the limits of wrapper libraries. Within 30 days of standard RAG on 100+ page insurance documents, I diagnosed catastrophic context collapse and citation failure. The fix wasn't a bigger context window — it was a custom runtime schema transpilation layer and a hierarchical node retrieval engine, built from scratch.
“The best GenAI architecture is the one that can provably tell you exactly why it gave the answer it did, every single time.”
FEATURED WORK
Production systems. Published research. Shipped products.
Lead Data Scientist · 2024–Present · Confidential Employer
Problem. Off-the-shelf LangChain RAG failed catastrophically on 700+ page insurance policies — hallucinating citations with legal consequences.
Build. Custom hierarchical JSON-tree retrieval with runtime schema transpilation. Deterministic citation validation streams output character-by-character. New carriers onboard via config — zero code changes.
XGBoost pipeline for claim triage. SHAP for 100% regulatory audit trail. 45% cost reduction · $200K annual savings · 35% faster settlements.
XGBoost · SHAP · Django · RabbitMQ
Team lead of 10. Fine-tuned Rat-SQL transformer. NL query → SQL → Plotly dashboard, self-serve. 87% accuracy on complex multi-table JOINs — before LLMs existed.
Rat-SQL · Fine-tuning · Plotly · Postgres
3-service microarchitecture (FastAPI + FastMCP + Azure Service Bus). PydanticAI agent generates type-safe models from per-tenant schemas at runtime. 95% alignment · 2% hallucination.
PydanticAI · FastMCP · Multi-Tenant
Multi-objective autoencoders for Brain-Computer Interfaces. Selective anonymization preserves clinical signal while eliminating re-identification. PST 2023, Copenhagen.
View on IEEE Xplore ↗SIDE PROJECTS
Browser-native AI video editor. Natural language → FFmpeg.wasm. TensorFlow.js face detection, all client-side. No uploads, no server, no privacy tradeoff.
React 19 · FFmpeg.wasm · TensorFlow.js · Vite
GitHub ↗Export ChatGPT, Claude, and Gemini conversations as TXT, Markdown, JSON, or HTML. Zero permissions, all processing local. Chrome + Firefox.
Chrome/Firefox Extension · Manifest V3 · Zero Permissions
GitHub ↗BY THE NUMBERS
Every number is earned, not estimated from a demo.
TECHNICAL ARSENAL
Machine Learning · Deep Learning · Generative AI · Agentic AI · Large Language Models · Computer Vision · Transformers · Explainable AI (SHAP)
FastMCP · PydanticAI · Hierarchical Node Retrieval · Citation Validation · Azure OpenAI / GPT-4o · Cohere Reranking · Embedding Models
PyTorch · Transformers · XGBoost · SHAP · Vision OCR · NLP / Fine-tuning · Prophet · Scikit-learn · Plotly · Matplotlib · Seaborn
Python · SQL
PySpark · Microsoft Azure (Cosmos DB IVF · Service Bus) · Google Cloud Platform · AWS (Lambda · SageMaker · S3) · RabbitMQ · PostgreSQL · MongoDB · Docker
FastAPI · OpenTelemetry · Arize Phoenix · Django · Flask · Adobe PDF Services · Team Leadership (10–14)
SIX YEARS · US · UK · AU · EU
CAREER TIMELINE
Primus Software Corporation · Waterloo, ON
Led cross-functional team of 10–14. Scaled enterprise AI to 2,000+ users globally. Resolved two production crises. Built zero-code carrier onboarding. Promoted Senior → Lead in 12 months.
Primus Software Corporation · Waterloo, ON
Diagnosed LangChain's fundamental limits on multi-document policies. Designed hierarchical RAG architecture solo in 3 months. Latency crisis: 2.5 min → 40 sec.
Lakehead University · Thunder Bay, ON
Project-based Masters, supervised by Dr. Garima Bajwa. Published privacy-preserving ML research at PST 2023, Copenhagen. Continued AI development at Primus concurrently.
Ciena · Ottawa, ON
PySpark pipelines, divisive clustering, and manufacturing batch anomaly detection.
Primus · Noida, India → Canada (2021)
Built FNOL classification for Crawford & Company solo. Led PriML NL-to-SQL project (team of 10). CTO recognition + bonus. Six years of insurance domain expertise starts here.
PUBLICATIONS
PST 2023 · Copenhagen, Denmark
Autoencoder architectures for securing biological telemetry — preserving clinical signal while eliminating re-identification vectors. Supervised by Dr. Garima Bajwa.
View on IEEE Xplore ↗ICICT 2019 · Springer
Optimized edge-compute inference for computer vision on resource-constrained hardware. In-memory strategies significantly reduce latency for real-time face recognition.
View on Springer ↗LEADERSHIP & COMMUNICATION
Data scientists who can communicate build better systems. The evidence:
Competitive public speaking that directly informs how I present technical findings to non-technical stakeholders — executives, clients, and insurance carriers.
Ran day-to-day technical and delivery decisions for a team of 10–14. Direct stakeholder requirement gathering, refinement, and brainstorming. Second-most senior on the team.
OPEN TO $190K+ LEAD & STAFF DS ROLES · KITCHENER, ONTARIO OR REMOTE
Big Tech, AI-native, or Enterprise AI. If your engineering bar is high and you need an architect who thinks in systems, let's talk.
GIRIJESH SINGH · LEAD DATA SCIENTIST · KITCHENER, ON · 2026