Markham, Ontario, Canada
Aatif Usmani
AI Engineer at Tasteport, one of the first 10 members of a VC-backed startup. I build AI agents, assistants, and RAG systems in production.
▸ press start · choose your class ◂
AI Engineer at Tasteport, one of the first 10 members of a VC-backed startup. I build AI agents, assistants, and RAG systems in production.
▸ press start · choose your class ◂
I build AI products at Tasteport, a VC-backed ethnic grocery platform where I was one of the first 10 employees. My day-to-day is architecting Mina, our production AI shopping assistant: RAG, tool calling, vector search, long-term memory, and the cloud infra that keeps it fast.
Outside work I do hackathons and other projects that I find interesting (Invisible City - 1st place YorkU Data Storytelling Challenge).
> init aatif_usmani --role
AI Engineer
Tasteport · Toronto, ON
6+
Years at Tasteport
location ... Markham, ON
email ...... ventures.aatif@gmail.com
github .... /AatifUsmani
education . B.Sc. Biochemistry, CS minor · transitioned to full-time at Tasteport
Production AI shopping assistant: Python, SQL. LangChain/Langgraph orchestration, PostgreSQL, GPT-4o, GPT-3.5-turbo, OpenAI embeddings, Pinecone vector database, long-term memory, guardrails (Pydantic validation, RAG, etc.) and Redis semantic caching on AWS Lambda. Langsmith + Cloudwatch. Under 3s p95 latency.
Autonomous self-healing data agent: detects schema drift, computes column-level blast radius from lineage, and uses deterministic YAML policy (not the LLM) to choose automated repair vs human approval, verified by static checks and a test eval harness.
Geospatial platform scoring Toronto facilities via toxicity-weighted exposure, Haversine proximity decay, and a four-detector anomaly ensemble combining Isolation Forests with interpretable domain rules. 1st place at the YorkU Data Competition.
Quick prototype I built for work and learned a ton from: started as brute-force embedding search over cached responses, then rebuilt on Redis vector search in production, skipping redundant LLM calls and cutting API costs at scale.
CX Intern
B2B, B2C & fulfillment relationships
Data Associate
Inventory data, ops tooling & audits
AI Engineer
Production AI agents & assistants
Architecting Mina, our production AI shopping assistant
See projects ^ for more info - All of the following (tool calling, memory, etc.) is implemented:
Tool calling for real-time product search
Semantic vector search over a live 10K+ SKU inventory using cosine similarity.
Long-term memory across sessions
Preferences auto-extracted from conversations with GPT-3.5-turbo (90% cheaper than GPT-4o), stored as embeddings in Pinecone for personalized recommendations.
Production-grade AI deployments
AWS Lambda + Docker with auto-scaling, CloudWatch monitoring, and API Gateway rate limiting for production traffic.
HNSW vector indexes
Sub-100ms queries over 25K+ embeddings across the product catalog and user memory stores.
Semantic caching in production
Redis + cosine similarity (>0.70) eliminates redundant LLM calls: 42% cache hit rate and 40% lower API costs.
AI / ML
Lv.87
EXP 119190 / 99999
RAG & Agents
Lv.84
EXP 115080 / 99999
Backend & Data
Lv.78
EXP 106860 / 99999
Cloud & DevOps
Lv.74
EXP 101380 / 99999
York University
B.Sc. Biochemistry, Minor in CS · transitioned to full-time work (2025)
AI & Data Society
IT Engineer (2025)
Wilfrid Laurier
Business & Science (2022–2024)
YorkU Data Competition
1st place - Invisible City
NVIDIA
Certified Associate - Generative AI & LLMs
DeepLearning.AI
Deep Learning Specialization