Shahroz Ahmad's Resume

Shahroz Ahmad

Full stack AI/ML engineer vibe—coding AI verticals and meticulous slops.

Halifax, Canada, AST

Shahroz Ahmad's profile picture

About

AI/ML engineer blending TypeScript product work with Python, LLM, and computer-vision pipelines. Ship production RAG automation, carbon intelligence analytics, and privacy-safe CV tooling on AWS Bedrock, Dagster, and cloud-native stacks.

Work Experience

Davies North AmericaFull-time

Full-Stack AI/ML Engineer

HalifaxHybrid
  • Leading automation of insurance claims with a RAG system on AWS Bedrock, LangChain, LlamaIndex, Textract, S3, and Lambda, replacing manual OnBase workflows.
  • Building a fraud detection app that connects React frontends, Python ML pipelines, and Dagster to analyze loss data within the ICM ecosystem.
  • Implementing MCP and A2A automations to cross-verify claims against descriptions and photos while integrating modern ML services with legacy OnBase systems.
  • Replaced manual OnBase claim reviews with AWS Bedrock + LangChain RAG flows.
  • Linked React apps, Python models, and Dagster fraud pipelines across ICM data.
  • Shipped MCP/A2A checks that compare claim text and photos across legacy stacks.
  • AI/ML
  • Amazon Bedrock

AcuicyContract

Machine Learning Researcher - LLM

HalifaxHybrid
  • Enhanced the NetZero engine by adding non-linear ML models tracked and deployed via MLflow, improving CAPEX, ROI, and emissions estimates by 20%.
  • Automated carbon data retrieval and fine-tuned LLMs with Adaptive RAG, LoRA, and QLoRA, serving them through vLLM for batched inference.
  • Built Dagster ETL pipelines that pull graph data from ArangoDB, process it, and land curated features in ClickHouse for training.
  • Connected ClickHouse to Superset for interactive analytics and shipped an open-source Superset Python client for microservice automation.
  • Lifted NetZero forecasts 20% through MLflow-tracked nonlinear models.
  • Automated carbon data ingestion and fine-tuned Adaptive RAG + LoRA LLMs on vLLM.
  • Built Dagster → ClickHouse → Superset loop for graph features and analytics.
  • Deep Learning
  • MLflow

DetectCo-op

Machine Learning Engineer - Computer Vision

HalifaxHybrid
  • Built a high-throughput EgoBlur inference service to anonymize faces and plates prior to delivering annotated inspection datasets.
  • Fine-tuned Faster R-CNN, YOLO, and Mask R-CNN models in PyTorch, pushing mAP beyond 0.7 for defect-feature pair detection.
  • Developed Dagster-driven pipelines that trigger training or retraining when new data or class labels arrive, using assets, partitions, and ops graphs.
  • Integrated MLflow and Hydra into the MLOps toolchain for experiment tracking, configuration, evaluation, and deployment.
  • Launched EgoBlur to anonymize inspection photos before delivery.
  • Tuned YOLO, Faster/Mask R-CNN models to exceed 0.7 mAP on defect pairs.
  • Drove Dagster retrains with MLflow + Hydra whenever new labels or data landed.
  • Deep Learning
  • MLflow

AcuicyContractPart-time

Machine Learning Engineer

HalifaxHybrid
Developed Amazon Bedrock powered services that automated client onboarding checks and accelerated proof-of-value engagements.
  • Built Bedrock copilots that automate onboarding diligence for clients.
  • Accelerated POV delivery by wiring Bedrock flows into compliance workflows.
  • Amazon Bedrock

Scale AIContract

LLM Engineer - Training

San FranciscoRemote
  • Produced high-quality training data across multiple programming languages and frameworks as a "Platinum" rank team member, leading 5+ campaigns that improved SOTA LLM model capabilities by 30%.
  • Audited training data, developed eval sets, and optimized model performance with RLHF by enhancing correctness, informativeness, clarity, and creativity, resulting in a significant LLM reasoning uplift.
  • Implemented chain-of-thought prompting techniques to improve the model's coding and reasoning abilities, achieving a 15% increase in problem-solving accuracy.
  • Enhanced SOTA LLM model performance by 40% on the SWE-bench dataset, improving code generation, bug fixing, and code documentation tasks.
  • Led 5+ Platinum data campaigns, boosting frontier LLM quality by ~30%.
  • Audited RLHF datasets and evals to raise SWE-bench scores 40%.
  • Applied chain-of-thought prompting to lift coding accuracy 15%.
  • LLM
  • RLHF

ArbisoftFull-time

Software Engineer

McKinneyRemote
  • Worked as an open-source core contributor in the Open edX community, helping revamp the "edx-platform" from a monolithic architecture to a distributed microfrontends and microservices architecture.
  • Reduced critical production and security issues by 20% for thousands of online learners through dependency upgrades and fixes.
  • Enhanced CI/CD pipelines with GitHub Actions, automating tasks like semantic versioning and repository translations, saving significant manual effort.
  • Developed a scalable web application for "Unlisted," handling thousands of concurrent users, scraping and indexing tens of thousands of property data points, and implementing 20+ AI-powered property proposal features.
  • Led quality engineering and data analysis on 50+ property features for "Unlisted," improving search criteria by 40%.
  • Core contributor migrating Open edX from monolith to MFE/services.
  • Cut prod/security incidents 20% through upgrades and automated CI/CD.
  • Built “Unlisted” real-estate platform powering 20+ AI proposal features.
  • Open edX
  • Microfrontends
  • Open-Source Contribution

Dubizzle LabsFull-time

Software Engineer

LahoreOn-site
  • Conducted in-depth research and integrated the ELK stack into the existing Propforce backend architecture.
  • Optimized spatial database indexing and implemented Elasticsearch geo-queries, reducing full-length address lookup times from 2-3 seconds to under 300 milliseconds.
  • Developed a scalable, multi-tenant backend for Propforms, a national land balloting project, supporting 1,000+ tenants and handling 10,000+ ballots daily with high availability.
  • Embedded ELK observability inside Propforce’s backend.
  • Cut geo lookup latency from seconds to <300 ms via spatial indexes + Elasticsearch.
  • Launched Propforms multi-tenant backend serving 1k+ tenants and 10k ballots/day.
  • ELK
  • Geo

i2cFull-time

Software Engineer

Redwood CityRemote
  • Developed globally active, highly scalable, multi-threaded digital payment backend services and batch schedulers for major clients, including CIBC, Sightline, Petal, and Vantage Bank.
  • Horizontally scaled the Direct Deposit Scheduler, increasing transactions per second from 50 to 500.
  • Improved customer care service by automating call evaluation with NLP, resulting in a 15% increase in the Customer Satisfaction Score.
  • Mentored junior engineers through training sessions and code reviews, leading to a 40% improvement in code quality and a 90% job satisfaction rate.
  • Shipped payment services for CIBC, Sightline, Petal, and Vantage Bank.
  • Scaled Direct Deposit throughput 10× (50 → 500 TPS) with new scheduler.
  • Automated NLP call scoring to raise CSAT 15% while mentoring juniors.
  • Payments
  • NLP

Education

Dalhousie University

Master of Applied Computer Science (MACS), Computer Science

National University of Computer and Emerging Sciences

Bachelor of Science (BS), Computer Science

Skills

Programming Languages

  • Python
  • TypeScript
  • JavaScript
  • SQL

Frameworks & Platforms

  • Django
  • Django REST Framework
  • FastAPI
  • GraphQL
  • REST API
  • Nginx
  • RabbitMQ
  • Keycloak
  • React
  • ShadCN
  • Mantine
  • TailwindCSS
  • Redux

Cloud Platforms

  • AWS
  • Azure

Side projects

gtdustIN PROGRESS

GTD desktop client that recreates Things 3 capture, project, and inbox flows on Windows using Rust and iced.rs for a lightweight, fast UI.

  • Rust
  • iced.rs
  • Cross-platform
  • Desktop

SkilledInDEPRECATED

Skills intelligence platform pairing a LinkedIn-scraping Chrome extension with an AI backend that extracts skills, normalizes job titles, and powers future analytics dashboards.

  • Chrome Extension
  • WXT