Ali Sajid

Assisted in building a platform that generates and calibrates software tasks for AI training. The challenge was making tasks difficult enough to be useful while keeping success objectively verifiable.

My contributions included reusable task libraries, calibration results, deterministic grading, worker routing, cancellation, and access controls. I also investigated agent runs that stalled after terminal errors and improved stop propagation, cleanup, and diagnostics. Tasks run in reproducible environments and are judged by concrete checks, rather than by whether an answer sounds convincing. The video walks through the task-calibration workflow.

Built an AI agent that reviews production pull requests and produces findings tied to specific files and lines. The workflow checks out the repository into a disposable workspace, allows the agent to inspect code using tools, and validates its structured findings before publishing.

I built the review workflow, GitHub integration, and safeguards with duplicate findings being filtered, file and line references being checked, and publication being blocked if the pull request changes during the run.

The application uses TypeScript, Fastify, React, SQLite, and the Cursor SDK.

Developed assessment features that turn project instructions and reviewer feedback into targeted quizzes. The system generates questions, grades answers, and controls when explanations and answers become available.

My work covered OpenAI integration, structured response validation, recovery from malformed model output, retries, adversarial filtering, and retake controls. I worked across the Python/FastAPI back end, React interface, and PostgreSQL data layer.

The focus was on making the assessment workflow dependable when model responses were incomplete or inconsistent.

Worked on an Insurify client engagement, first as a full-stack engineer and later as an engineering manager leading 10 engineers.

My work included migrating Backbone interfaces to React, moving WordPress sites to Gatsby, and building Next.js verticals such as Compare Finance. I also contributed Python microservices and partner APIs, integrated Braze and Klaviyo retention workflows, and implemented consent synchronization and analytics using GTM, gtag, and advertising pixels.

Alongside implementation, I reviewed designs and code and mentored engineers.

Worked on a CV extraction tool that turns unstructured resumes into information a reviewer can inspect and organize.

My contributions included domain taxonomy, multi-sheet views, saved views, and deployment work. The workflow makes extraction failures explicit instead of silently treating missing data as a valid result. I focused on making the output useful for day-to-day review by using consistent categories, practical filtering, and views users can return to.

The video demonstrates the candidate-review workflow.

A generated test is only useful if it catches the original failure and accepts a correct solution. I contributed to a verification pipeline that runs coding tasks in Docker and checks that distinction.

My work covered verifier execution, concurrency controls, resource logging, and retry handling. The pipeline checks baseline failures and reference-solution passes rather than trusting a model's explanation of success.

This gave reviewers concrete evidence of execution when assessing task quality and diagnosing unreliable runs.

Built accounting operations software that brings client records, document processing, and transaction review into one workflow.

My work covered Django/React features, permissions, activity tracking, OCR-based transaction extraction, OpenAI-assisted categorization, and Excel exports. Background jobs use Celery and Redis, with integrations including Microsoft Graph, Stripe, and Twilio. I built the review steps so users can inspect and correct extracted information before using it in their accounting work.

This was part-time consulting work for Yogi & Associates.