ALL PROJECTS

Problem: Turning raw screencast or facecam footage into a usable rough cut means manually scrubbing through hours of audio to cut stutters, dead air, and rambling takes — tedious work that delays the creative editing that actually matters, and most AI editing tools lock footage into a cloud pipeline or a proprietary subscription instead of a professional NLE workflow.

Result: Built a full desktop application that transcribes footage, uses an LLM to make AI-driven keep/cut decisions, and exports either directly or as a fully editable OpenTimelineIO timeline for DaVinci Resolve — with a custom LCS-based text-to-timestamp aligner and waveform-aware boundary snapping so cuts land on natural pauses rather than mid-word or mid-syllable.

Problem: Editing talking-head and tutorial videos manually is repetitive and slow, especially when removing retakes, filler words, and pauses before creating a first cut.

Result: Built an end-to-end AI-assisted editing pipeline that automates transcript-based rough cuts, supports human-in-the-loop review, and exports directly to production formats (MP4 and DaVinci-compatible timelines).

Problem: Modern hiring platforms can't verify that the person on screen during a video interview is actually the invited candidate — they may join under a pseudonym, a device name, or a mismatched display name. A one-shot name match is trivially defeated; the system needs to reason continuously as new signals arrive throughout the call.

Result: Built a complete three-service system (simulator, Next.js dashboard, Python inference engine) that runs 9 independent signal extractors in parallel — 7 rule-based and 2 LLM-backed — and maintains a live Bayesian belief over candidate identity, surfacing a real-time verdict with confidence and a per-signal reasoning trail on the dashboard.

Problem: Building a robust personal AI assistant requires integrating multiple components: LLM providers, communication channels, tool execution, and maintaining conversation context, all while ensuring security, extensibility, and performance.

Result: Developed a fully functional AI assistant with 8 built-in tools, supporting multiple LLM providers and Telegram integration, demonstrating expertise in async Python development, API design, and agent-based architectures.

Problem: Official Reddit clients are heavy, ad-laden, and offer little control over feed behavior, image quality, or session persistence; there was no lightweight, cross-platform Reddit client that handled infinite scroll, gallery posts, and search correctly while staying fast on mobile data.

Result: Built and shipped a cross-platform (Android/iOS/desktop/web) Flutter Reddit client with cookie-based auth, true infinite-scroll feeds, site-wide search with sort-by, and swipeable gallery posts; a later audit pass found and fixed 10+ concrete bugs including a broken pagination model that was losing already-loaded posts, a dropped session-cookie rotation bug, N+1 author-avatar requests on every feed load, and full-resolution image decoding that was inflating memory use well beyond what was ever rendered on screen.

Problem: Large screenshot folders become hard to search and maintain. Manual sorting is repetitive, inconsistent, and usually skips useful metadata like what the screenshot contains.

Result: Built a working AI-powered CLI that auto-categorizes screenshots, renames/moves them into structured folders, stores searchable metadata in SQLite, and generates an interactive HTML report for quick review.

Problem: LLM agents often receive full schemas for all available tools on every turn, which bloats context, increases token cost, and hurts tool-selection accuracy.

Result: Built and published a Python package on PyPI that introduces a 3-meta-tool execution layer, reducing tool-context overhead by exposing tools on demand and using Python function-call syntax with AST-based validation. Added robust validation/error formatting and 90 automated tests.

Problem: Creating short-form videos from mixed assets (images, clips, subtitles, transitions) usually requires heavy editors or brittle scripts, making repeatable content production slow and hard to automate.

Result: Built and published a reusable Python library on PyPI (v0.1.4) that renders scripted videos through a deterministic JSON project format, with modular effects/transitions, subtitle animations, and configurable resize strategies.

Problem: Critical alerts can be missed when users rely on passive notifications. OpenAlert was built to deliver high-priority, actionable mobile alerts and connect external events (like inbox webhooks) to immediate user attention.

Result: Delivered a working full-stack alerting system with authenticated mobile users, device token management, webhook-driven push notifications, and in-app alarm controls. The project demonstrates end-to-end ownership across mobile client, API design, authentication, notification delivery, and deployment-ready configuration.

Problem: Voice assistants lack real-time, full-duplex conversation with contextual reasoning and dynamic tool use.

Result: Built a real-time voice assistant with on-device VAD, STT, and fast TTS — enabling natural, low-latency conversations with LLM-powered tool use.

Problem: Creating polished video content from text stories requires hours of manual editing and production work.

Result: Engineered a complete AI-powered video automation pipeline, reducing production time from hours to under 2 minutes with 15x faster rendering than MoviePy.

Problem: People preparing for interviews and presentations have no way to get detailed, AI-driven feedback on their speech quality.

Result: Built a full-stack generative AI platform analyzing 1,000+ speeches with 25%+ better feedback accuracy using Whisper and GPT-4.