Safi Ullah

A browser-based 2D momentum skating game built as an interactive portfolio experience. It combines Canvas rendering, procedural terrain, risk/reward route design, power-ups, mobile controls, sound design, and a polished Techno-Oasis visual direction.

• Built a custom delta-time game loop without a game engine
• Implemented momentum-based skating, jumping, pumping, and speed-sensitive controls
• Designed procedural terrain with risk/reward routes, recovery sections, and score-based difficulty
• Added power-up systems: Overclock, Focus Mode, Score Surge, and Patch Pulse
• Shipped responsive desktop and mobile gameplay with landscape controls, audio feedback, and Techno-Oasis art direction

A distributed content intelligence pipeline that automatically discovers technical signals, synthesizes them into narrative drafts using AI, and publishes approved content to social platforms. Five decoupled microservices handle ingestion, transcription, synthesis, delivery, and orchestration independently — so failures in one layer never cascade to others.

• Five-service event-driven architecture — feed-ingestor, transcriber, content-brain, publisher, and orchestrator — where no service has direct knowledge of another
• AI at the ingestion layer: Zod-validated structured outputs force the LLM to classify and summarize raw feeds before they reach the database, treating AI as a deterministic pipeline stage
• Deepgram speech-to-text pipeline converts voice memos directly into transcripts that enter the synthesis queue, bridging unstructured physical capture and the structured content pipeline
• Fault-isolated publisher with zero AI dependencies — external API calls to X and LinkedIn are fully separated from core intelligence, enabling independent retry queues and rate-limit handling
• Pure event-router orchestrator with no database or external API dependencies — listens for events, sequences jobs, and keeps all other services completely decoupled

A skill calibration engine that trains learners on targeted micro-skills through adaptive difficulty. Problems are served based on a live performance rating derived from an append-only attempt history, designed to keep learners at their exact edge of competence.

• Append-only attempt system — skill rating is derived and recomputed from full history, no mutable state
• Clean Architecture enforced across all layers: UI → API → Domain → Data, with the domain layer forbidden from touching Prisma or environment variables
• Adaptive difficulty calibration that always serves problems at the learner's current performance edge
• Micro-interaction design following the Unmistakable Joy Doctrine — kinetic hover states, streak heat systems, and answer-reveal delays engineered for flow
• Strict TypeScript throughout with no any allowed; Zod used to make illegal API states unrepresentable

A Chrome extension that acts as a client-side firewall against Reddit's algorithmic attention mechanics. It restricts navigation to user-defined subreddits and surgically removes infinite feeds, sticky sidebars, and right-rail recommendations — turning Reddit into a focused reference tool.

• MutationObserver with batched flushing (MAX_BATCH_NODES = 60, FLUSH_DELAY_MS = 120ms) to sanitize Reddit's SPA DOM changes without blocking the main thread
• Capture-phase event listeners that preempt Reddit's React routing before it processes navigation clicks
• Spatial DOM heuristics using getBoundingClientRect() to identify and strip the right rail — resilient to Reddit's frequent class-name changes
• History API hooks (pushState / replaceState) enforcing subreddit boundaries across Reddit's client-side routing
• Immediate rule enforcement before chrome.storage.sync resolves, eliminating any flash of unrestricted content

A Chrome extension that strips Instagram's algorithmic content and enforces a chronological, followed-only feed. It operates across three interception layers — network (GraphQL payload scrubbing), DOM (MutationObserver heuristics), and navigation (History API patching) — to surgically remove Reels, Explore, Suggested Posts, and Sponsored Content.

• Three-layer defense: intercepts GraphQL responses before they reach the DOM, removes injected elements via MutationObserver heuristics, and blocks SPA navigation to restricted routes
• Page context injection — overrides window.fetch and XMLHttpRequest inside Instagram's main world to bypass the Content Script isolated environment
• Algorithmic learning mechanism that silently parses GraphQL metadata as the user scrolls, building a persisted cache of followed accounts to improve filter accuracy over time
• Secure cross-context messaging pipeline between the isolated content script, injected page context, and extension popup via postMessage and chrome.runtime
• Debounced DOM observation to prevent performance bottlenecks from Instagram's frequent React re-renders