- Device Enumeration & BAR Decoding: Simulated PCIe bus scans, Base Address Register (BAR) mappings, and dynamic discovery of CXL-compliant devices.
- Coherency & Protocol Layers: Implemented cache-coherency state machines, host↔device messaging, and basic memory access flows.
- MCTP Integration: Added Memory Component Transport Protocol support for robust in-band device control and error reporting.
- External ROM & Memory Ops: Modeled ROM fetches for firmware-level initialization and sketched host-managed memory pooling scenarios.
- Ongoing Enhancements: Extending dynamic device hot-plug, advanced MCTP message queues, and validation against Intel/CXL Consortium reference flows.
- Architecture & Goal: Designed a plugin (in Windows) and systemd service (in Linux) for semantic image search using multimodal AI techniques, enabling natural language and visual similarity queries over large image datasets.
- Text-to-Image Retrieval: Integrated OpenAI’s CLIP (
clip-vit-base-patch32) model to map text prompts into visual embedding space and retrieve semantically relevant images. - Image-to-Image Matching: Enabled visual similarity search by comparing vision embeddings to find near-duplicate or thematically similar content.
- Embedded Text Indexing: Incorporated Doctr OCR to detect and index text inside images for searching screenshots, scanned documents, and meme captions.
- Minimal & Fast UI: Built a clean, responsive interface supporting directory-level image indexing, dynamic threshold tuning, and ranked top-K results.
- Custom Ranking Controls: Allowed user-defined similarity thresholds and output sizes to fine-tune retrieval precision and recall.
- Backend & Data: Architected a Supabase (PostgreSQL) backend with row-level security, hosting 500+ curated food–drug interaction records.
- Search & Caching: Engineered full-text search indices and in-memory caches—reduced average lookup latency by 40%.
- Production Impact: Live at Deenanath Mangeshkar Hospital since Apr 2025—empowering HCPs with instant safety alerts.
- Excel Automation: Automated student roster import/export and attendance marking in `.xlsx` via `openpyxl`.
- Real-Time Sync: Persisted logs to Firebase Firestore, enabling live updates across teacher dashboards.
- Reliability & UX: Added retry logic for recognition failures and user prompts—achieved 98% match accuracy in mixed-lighting conditions.
📧 karunyachavan84@outlook.com | 📞 +91 93592 88942 | 🔗 LinkedIn | 🔗 GitHub