THE ML ENGINEER'S
QUEST
an interactive resume by Sahand Somi — Machine Learning / AI Engineer
or just scroll — every world loads either way
Edmonton, AB · ssomi@ualberta.ca · (587) 988-8682
The Foundry of Origins
Before cloud pipelines and AI centres of excellence, there was a research assistant at the University of Alberta, buried in construction schedules and hand-classified drawings. The boss here is built from every hour of manual scheduling — slow, heavy, and immovable.
VICTORY
- Designed a hybrid Agent-Based Modeling + Reinforcement Learning framework for construction resource scheduling — cut project time 20% (published, Automation in Construction)
- Automated a manual drawing-classification task, cutting effort 40%
ORIGIN STATS
- M.Sc., Construction Management & Engineering — University of Alberta (2018–2020), GPA 4.0/4.0, focus: AI Applications in Construction
- B.Sc., Civil Engineering — University of Tabriz, Iran (2013–2017), ranked 1st in class
- NSERC Canada Graduate Scholarship earned during the Master's program
Startup Swamp
Co-founding Bawse meant ten teammates, one MVP deadline, and a swamp of unstructured survey responses. The Chaos Imp feeds on scope creep — the only way through is fast, scrappy NLP and a team that ships.
VICTORY
- Led a team of 10 building an MVP mobile app (Firebase, Node.js)
- Used NLP (NLTK) to extract insights from 200+ survey responses
- Segmented 100+ customers from the results
The Mining Depths
At AltaML, a mining client needed more yield from the same ore. The Yield Ogre guards the optimal input mix, deep inside a simulated production process. It only falls to evolution.
VICTORY
- Modeled the production process as a simulation and applied a genetic-algorithm optimizer to find the yield-maximizing input mix — beat the benchmark by 20%
- Built a data pipeline to clean and prepare 2M+ data points (Azure ML, Blob Storage)
- Scoped the problem directly with stakeholders and led a 2-engineer team
Frostlands of Forecasting
At Amii, an ice-slow training pipeline threatened to freeze an 800-product forecasting engagement solid. The Slowtrain Wyrm took two hours to move an inch — until multiprocessing cut it down to fifteen minutes.
VICTORY
- Partnered with AWS to build sales-forecasting models for 800 warehouse products (SageMaker, Athena, SQL)
- Built a multiprocessing training pipeline cutting training time from 2 hours to 15 minutes (8×); added a feature store + MLflow tracking
- Created XGSleeve, an XGBoost model detecting oil-well incidents from sensor data, deployed at the edge with Docker
- Improved a time-series classification model 30% (Azure Cosmos DB, Azure ML); built pipelines pulling 3M+ sensor data points
The Municipal Fortress
As Senior Data Scientist, Sahand leads a team of 5 junior engineers and works alongside 5 senior engineers, 10 data engineers, and a solution architect across the City of Edmonton's AI Centre of Excellence — promoted to the top of the technical band. The fortress has four strongholds. Each has its own boss.
VICTORY
- Replaced physical traffic sensors with an event-driven GCP pipeline (Pub/Sub → Cloud Functions → Dataflow/Apache Beam) streaming camera frames into BigQuery
- CV inference deployed on GKE via a sidecar pattern, decoupled from the app for independent retraining; derives vehicle speed from monocular frame sequences via per-camera calibration
- Saves the City roughly $500K/year versus physical sensor hardware
VICTORY
- Built a generalized internal agent library on Google's Agent ADK — data-connection agents, a semantic registry, and a tool-calling framework
- Packages outputs as ready-to-use agent functions, plus dynamic UI generation for natural-language requests
- Scaled to 5 teams across the transit department, removing the need for a dedicated analyst on ad hoc requests
VICTORY
- Automated ingestion of corporate safety incidents into SAP
- An LLM extracts structured fields (injury source, classification) from free-text reports, surfaced through an advisor-review dashboard
- Cuts safety-advisor review time 90%/week while keeping a human in the loop on safety-critical data
VICTORY
- Forecasts public-safety incidents across 20,000 geographic grids using 2M+ records pulled from 10 source systems with inconsistent call-type naming
- A funnel resolves the obvious majority with cheap keyword/string-similarity matching, escalating only ambiguous cases to an LLM
- Normalizes taxonomy across all 10 sources while cutting matching cost
GEAR UPGRADES EARNED (Platform & Standards)
- Led the migration from an R-Shiny monolith to a Flask/React sidecar architecture; added Prometheus/Grafana observability — ~90% reduction in average app startup latency (to ~2s), pipeline uptime up to 99%
- Standardized project scaffolding via Cookiecutter/Cruft across 40+ jobs and 30+ apps; hardened (DHI) base images cut detected vulnerabilities ~99%; adopted uv for Python dependency management
- Introduced a Gemini-based agent for merge-request review — average approval time cut from 2 days to ~4 hours
- Built an automated bug pipeline (intake form → ClickUp → Gemini-suggested fix → MR) — bug-fix turnaround from ~1 day to ~2 hours
The Estimation Tower
Now Principal AI Engineer at Rotaflow, Sahand leads AI integration strategy across engineering and operations. The Takeoff Titan is built from years of manual sprinkler and underground-piping takeoffs — every rivet a code reference. It only falls to an LLM that knows the codebook by heart.
VICTORY
- Built an LLM-powered estimation application for sprinkler and underground piping takeoffs that applies relevant codes and standards to generate compliant estimates — cutting manual estimation time
- Mapped enterprise processes with RACI and SIPOC frameworks, then automated the highest-value steps using agentic workflows
Boss Rush Complete
★ ALL 6 BOSSES DEFEATED ★Final stats, unlocked items, and how to reach the player behind the pixels.
LANGUAGES
- Python
- TypeScript
- JavaScript
- Rust
- SQL
ML / AI
- PyTorch
- TensorFlow
- Scikit-Learn
- Hugging Face
- XGBoost
- LLM Agents · RAG
- Elasticsearch
AGENTIC SYSTEMS
- Google Agent ADK
- MCP
- Semantic Registries
- Multi-Agent Orchestration
- Google Workspace API
GCP & DATA INFRA
- Pub/Sub · Cloud Functions
- Dataflow (Apache Beam)
- BigQuery · GKE
- AWS (SageMaker, Athena)
- Azure (ML, Cosmos DB, Synapse, Functions)
FULL-STACK & INFRA
- React · Flask · Node.js
- Docker · Kubernetes
- Sidecar Architecture
- Cookiecutter/Cruft · uv
- DHI Hardened Images
ENGINEERING PRACTICES
- CI/CD (GitLab)
- Automated Testing
- Pre-commit Hooks
- LLM-Assisted Code Review (Gemini)
- Observability (Prometheus, Grafana)
DATABASES
- PostgreSQL · MySQL
- BigQuery · Cosmos DB
- NoSQL · Blob Storage
TROPHIES EARNED (Certifications)
ACHIEVEMENTS UNLOCKED (Awards)
SIDE QUESTS COMPLETED (Talks & Presentations)
- Public Safety and AI — Upper Bound Conference 2024, Edmonton, AB
- RL Competition — Upper Bound Conference 2024, Edmonton, AB
- AI in the Public Sector
- Scaling AI Teams
CONTINUE?
insert coin to load next opportunity
Edmonton, AB