archit@portfolio:~$ whoami
Archit Konde
Machine Learning Engineer
█
open to opportunities
archit@portfolio:~$ █
I'm a Machine Learning Engineer with an MEng in Electrical & Computer Engineering from the University of Windsor.
I build AI systems end-to-end and actually test them — from writing retrieval algorithms by hand to deploying production-ready pipelines. My current focus is on LLMs, RAG architectures, and the engineering that makes AI reliable in production.
Currently based in Waterloo, ON — open to ML/AI engineering and software engineering roles.
Fundamentals of Neural Networks
International Journal for Research in Applied Science & Engineering Technology (IJRASET)
- Authored a research paper examining the architectural and mathematical foundations of neural networks
- Covered activation functions, backpropagation, gradient descent, and network topologies
Data Analyst
City of Windsor — Affordable Transit Program
- Analyzed transit ridership data across multiple routes to surface usage patterns and support service planning decisions
- Cleaned and restructured raw operational datasets; delivered summary reports used by program coordinators for planning cycles
Electrical & Computer Engineering
University of Windsor — Canada
Computer Engineering
University of Mumbai — India
supportops_ai_monitor.py
Built an LLM-powered ticket triage system using GPT-4o-mini for automated classification and priority scoring via structured multi-step tool-calling. Conditional routing logic, SQLite persistence, observability dashboard. Docker deployment, CI passing.
rag_from_scratch.py
Complete RAG pipeline built in pure Python — custom chunker, Okapi BM25, NumPy vector store, hybrid retrieval via Reciprocal Rank Fusion, cross-encoder re-ranking. 87 unit tests. No LangChain, no LlamaIndex. Hybrid + Rerank achieved MRR 1.0.
triagegeist_solution.py
Emergency triage acuity prediction on 80k clinical ED records. Pushed accuracy from 0.891 to 0.9995 CV via TF-IDF scaling on chief complaint text + 3-tier hybrid: deterministic lookup (99.4%), glaucoma-specific binary classifier, LightGBM fallback. $10k Kaggle competition.
insurance_reshopping_predictor.py
Predicts whether you’d benefit from re-shopping your car insurance — ML trained on 381K real insurance profiles. Data quality first: 8-check validation pipeline with SQL-style audit queries, LightGBM classification, SHAP waterfall explanations, counterfactual tips.
RAGOps API — Production RAG with FastAPI and pgvector
Upgrading a from-scratch RAG pipeline to a real API. Persistent vector storage with pgvector, layered FastAPI architecture, offline evaluation with Precision@k/Recall@k/MRR, and CI regression gating. The gap between notebook and production.
Insurance Re-Shopping Predictor — Data Quality First
Why data quality matters more than model accuracy in insurance ML. 8-check validation pipeline with SQL-style audit queries, LightGBM on 381K profiles, SHAP explainability, and honest limitations of training on Indian market data for North American predictions.
Building SupportOps AI Monitor — What I Learned
Architecture decisions, bugs found, deployment challenges, and honest answers to
questions I couldn’t answer at first. Covers applymap() removal in pandas 2.2, SQLite
ephemeral containers, Streamlit’s PWA limitations, and more.
Building RAG From Scratch — A Complete Technical Deep-Dive
Every algorithm in a RAG pipeline derived from first principles. BM25 with Robertson-Walker IDF, mean pooling math, cosine similarity as dot product, Reciprocal Rank Fusion, cross-encoder reranking, and evaluation metrics. No LangChain, no LlamaIndex.
Triagegeist — From 0.891 to 0.9995 CV Accuracy
How text beat everything else in emergency triage prediction. TF-IDF scaling experiments, error analysis tracing every mistake to a single diagnosis, and a 3-tier hybrid that routes 99.4% of predictions through a lookup table. $10k Kaggle competition.
I'm actively looking for opportunities in AI/ML engineering and research. If you're working on something interesting — let's talk.