Dual-Core ML Accelerator
Team INT8 attention accelerator with SRAM-backed MAC arrays, asynchronous sum exchange, and normalization. Compare the reported timing, power, and area tradeoffs.
INT8Async FIFOSRAMTeam project
Selected Work
Hardware, tested.
Tools, built to use.
Explore RTL systems, implementation results, and supporting software. Each project brings its case study, source, and reports together.
15 projects across hardware, tools, and software.
Hardware & Research
RTL implementation, verification, and measured engineering results.
SystemVerilog Viterbi Decoder
An 8-state decoder with BMC, ACS, banked survivor memory, and traceback. The report records 22 error-injection cases, including successful recovery and failure cases.
8 statesRate 1/222 error cases
Reported passes: 138/156 → 141/156
MAGE RTL Generation
Extended a multi-agent Verilog generation flow with failure routing, route-specific repair prompts, edit acceptance guards, and task-local debugging memory.
Verilog
MAGE
LLM Debugging
Icarus Verilog
Details
Problem
LLM-generated RTL can compile, simulate, and still fail from interface, timing, or behavioral mismatches.
Approach
Classified failures into syntax, interface, logic, or ambiguous routes and preserved candidate checkpoints per task.
Evidence
Routing cut estimated cost and runtime by 3.9%; memory raised full-set raw passes from 138/156 to 141/156.
Benchmarked on
VerilogEval-V2
Control layer
routing + local memory
The report frames routing as an efficiency control and memory persistence as the stronger correctness signal for future combined experiments.
Team report · GPU cryptography benchmarks
GPU-Accelerated GGM Tree Construction
Team CUDA cryptography project for GGM tree construction. My contribution focused on integrating Spongent and Keccak backends.
CUDA
GGM Tree
Keccak
Spongent
Details
Problem
GGM trees need millions of PRF evaluations, making throughput the bottleneck for protocol-scale workloads.
Approach
Used a flat breadth-first memory layout and one GPU thread per parent node to expand levels in parallel.
Evidence
Keccak reached 202M leaves/sec at depth 20; Spongent GPU time dropped from 2041 ms to 108 ms at depth 12 after optimizations.
Memory layout
flat BFS array
Main optimization
DRAM to registers
The report compares software-friendly Keccak with hardware-native Spongent, showing why relative speedup and absolute throughput tell different performance stories.
Engineering Tools
Interactive tools and native utilities.
Boolean logic → Verilog → CMOS
Logic & CMOS Studio
Browser tool for Boolean logic, truth tables, K-maps, simplified equations, Verilog, and static CMOS pull-up / pull-down schematics.
Digital Logic
CMOS
Verilog
K-map
Details
Problem
Logic classes often split truth tables, K-maps, Verilog, and transistor reasoning into separate tools.
Build
One browser studio connects Boolean input to simplified equations, generated HDL, and static CMOS views.
Signal
Shows hardware reasoning across abstraction layers instead of stopping at a polished interface screenshot.
Built around
logic-to-layout reasoning
Interface goal
compare views quickly
The project emphasizes traceability between Boolean input, minimized logic, generated Verilog, and static CMOS structure.
Structured practice for hardware interviews
Hardware Interview Trainer
Local-first interview practice app with a large question bank, mock rounds, review history, progress analytics, and JSON import / export.
Hardware Prep
Analytics
Local-first
Details
Problem
Hardware prep is hard to repeat when questions, missed concepts, and timing live in scattered notes.
Build
Timed sessions, review history, analytics, and import/export turn practice into a durable loop.
Signal
Local-first data keeps the tool useful across multiple sittings without making prep feel heavy.
Built around
repeatable practice loops
Data model
portable local progress
Question sessions, missed-question review, and analytics are designed to keep hardware interview prep structured over multiple sittings.
Native utility with privacy-first file handling
MarkMacZip
Archive utility for macOS that compresses and extracts common formats, supports password workflows, and keeps files on device.
macOS
Compression
Privacy
Details
Problem
Archive workflows are frequent enough to deserve a focused native utility instead of a loose script.
Build
Compression, extraction, passwords, job history, and releases are wrapped in a Mac-first interface.
Signal
The project balances product polish with privacy-first local file handling.
Built around
native file workflows
Product focus
local-only archive handling
The app frames compression, extraction, password handling, and release delivery as a polished desktop utility rather than a script wrapper.
Other Projects
Embedded systems, XR, maps, and data applications.
Coursework workflow
CourseWatch
Canvas deadline tracker for the macOS menu bar, built as a local-first coursework companion.
Details
- Keeps deadline checks close to the operating system instead of another browser tab.
Focused on compact status visibility, deadline scanning, and a native-feeling menu bar workflow.
Document privacy
MarkPDFLock
Password-protects PDFs on Mac with batch export and local-only document handling.
Details
- Frames a common document-security task as a fast native utility.
Built around a narrow document workflow: select PDFs, apply protection, and export without sending files elsewhere.
Private workflow assistant
JobAgentMac
Offline job-application assistant focused on keeping application workflow data on device.
Details
- Prioritizes local storage for sensitive job-search notes and status tracking.
Organizes application state, notes, and follow-up context with an offline-first privacy posture.
Embedded sensing
Indoor Environmental Monitoring System
ESP32 environmental monitor for CO2, comfort metrics, and live indoor status display.
Details
- Connects sensor data to clear room-status feedback and practical thresholds.
Pairs embedded sensing with a simple status interface so readings are useful at a glance.
Conversation-aware avatar prototype
CloserXR Sales Negotiator
Meta Quest mixed-reality prototype where a sales-agent avatar practices a life-insurance pitch through speech, gesture, and spatial movement.
Details
- Integrates speech input, Gemini conversation history, Android TTS, room-aware pacing, and a live VR status panel.
Routes user speech through AutoVAD or Android STT, maps conversation intent to avatar gestures, and mutes the mic during TTS to reduce echo pickup.
Geographic exploration
China Travel Map
Scenic-area map with search, language switching, images, and region highlights.
Details
- Combines location browsing, multilingual labels, and image-backed place context.
Uses map browsing and structured place metadata to make scenic-area exploration more visual and searchable.
Dataset browser
Google Maps Landmark
Static landmark map for the Google Maps miniatures list, using Leaflet for browsing and filtering.
Details
- Turns a landmark collection into a searchable, map-based browsing interface.
Transforms a static landmark list into an interface with geographic context, filtering, and fast inspection.
Model experimentation
Climate Emulation Model
Kaggle competition work around climate emulation modeling and machine learning experimentation.
Details
- Documents a model-development workflow around climate emulation data.
Captures experimentation around climate model emulation, data preparation, and notebook-based iteration.
No matching projects. Try another search or clear the filters.