Bibi
Say its name and it opens apps, moves the real mouse, reads the screen, and answers out loud. A custom agent loop routes each request across local models and Claude by task signal, cost, and latency.
The agent loop matters more than the model.
Built production RAG serving 10,000+ users at WellX AI. Co authored a benchmark paper accepted at ECCV 2026. Two local first products live on this domain. I look for the unglamorous fix first.
Gemini · MongoDB Atlas vector search · Flask · same pattern as my WellX AI work. No chat history is stored.
Everything below is clickable and verifiable. Live links, real repos, honest scope.
Say its name and it opens apps, moves the real mouse, reads the screen, and answers out loud. A custom agent loop routes each request across local models and Claude by task signal, cost, and latency.
The agent loop matters more than the model.
No bot joins the call. It records both sides at the OS audio layer and transcribes a 45 minute call in roughly 90 seconds on the Apple Neural Engine. Everything stays as plain files on your Mac.
Granola removed the bot from your meeting. MeetingScribe removes the cloud.
On paperCo author on a multimodal PCB schematic benchmark, arXiv:2604.00270.
Real storyI owned the human annotation pipeline: 1,854 real world schematics, about 109.9K labeled instances aligning 423.4K semantic labels to visual regions. I wrote the validity spec and enforced annotator consistency. Benchmark numbers only mean something if the ground truth is consistent.
Honest scopeMine: the annotation pipeline and its quality bar. Not mine: methodology design.
On paperA real time UAE news platform with LLM based source verification.
Real storyBuilt end to end with Claude Code in days, not weeks: Next.js, TypeScript, Tailwind, MCP server integrations. The launch post reached 26,000+ impressions on LinkedIn.
Honest scopeDeployed through Vercel's Git integration. No custom CI pipeline, and it does not need one.
On paperA SwiftUI app that scans grocery receipts and tracks expiry.
Real storyThe vision model kept hallucinating on POS abbreviations like BNLS SKNLS CKN BRST. Rather than throw a bigger model at it, I traced the failure and built a ReceiptNormalizer that expands abbreviations deterministically before the model ever sees the input. The unglamorous fix was the right one.
Honest scopeNot on the App Store. A working system I use, not a shipped product.
On paperAn encrypted file system in C.
Real storyAES-256 block encryption, SHA-256 integrity, PBKDF2 key derivation, crash recovery journaling, and a custom LRU cache that cut repeated file access latency 65%. I wanted to understand what actually happens at the block layer.
Honest scopeA learning build, and it taught me more about storage than any course did.
On paperA marketing site for a consulting firm.
Real storyHand coded solo in vanilla HTML, CSS, and JavaScript under a signed contractor agreement: shared component system, scroll reveals, an ambient canvas hero. Real client, real deadline, real invoice.
Honest scopeNo framework because the job did not need one.
Health tech startup, reported directly to the CTO. Built production RAG pipelines serving 10,000+ users over enterprise wearable data with Python, FastAPI, and vector search. Designed REST ingestion for Fitbit, Apple Watch, and WHOOP. Owned AI features from prompt design through offline evaluation, deployment, and monitoring.
Real storyA chunk of the job was tracing an outsourced production stack, documenting what was actually running, and fixing the parts with no visibility. The real bottleneck was not building new things. It was understanding what was already there.
Modernized legacy research IT services: migrated backends to C# .NET, wrote parameterized SQL to close injection holes, built CI/CD on Azure DevOps, containerized services with Docker. Delivered 11 backend enhancements and cut deployment time 40%.
Shipped a LangChain document Q&A and content generation system that handled 1,000+ production queries, integrating OpenAI and Cohere. Cut latency 40% with prompt chaining, chunking, and caching. Built a lightweight LLM eval harness and tuned outputs against direct researcher feedback.
Led BACnet integration on a real time HVAC simulation system: Python backend, React dashboard, historical weather playback so sponsor engineers could test behavior without touching live infrastructure. Coordinated weekly with the sponsor's team.
Built React UI components and customer facing booking flows on a travel platform.