A production-grade chunk-aware diffing engine for Retrieval-Augmented Generation systems, fused with a decentralized vector search framework. Eliminates redundant embedding API calls at scale while maintaining sub-50ms search latency across distributed sharded nodes. Features a React-powered visualization layer for real-time spatial and indexing observability.

Attribute	Details
Stack	Python · React · ChromaDB · gRPC · KD-Trees · SQLite · MapReduce · Multithreading
Scale	Multi-node sharded architecture · concurrent query fanouts · high-volume sync
Performance	>90% reduction in embedding API calls · sub-50ms search latency · O(1) retrieval targets
Concurrency	ThreadPoolExecutor + ChromaDB · zero race conditions · Leaky Bucket rate limiter
Algorithm	Hirschberg's space-improved LCS from scratch · hyperplane-distance pruning for KD-Trees
Repository

The core insight driving VectorDiff is that most RAG sync operations are redundant — documents change at the chunk level, not the document level. By implementing Hirschberg's LCS algorithm from scratch, the system performs surgical delta detection on chunk sequences, avoiding re-embedding unchanged content. The distributed search layer uses a custom gRPC MapReduce framework, fanning out queries to SQLite-sharded worker nodes and merging results through a custom KD-Tree accelerated by hyperplane-distance pruning that aggressively eliminates low-probability spatial regions before scoring.

A ground-up implementation of a production-grade distributed caching system in Java, designed to match the correctness guarantees of enterprise-grade systems like Redis. Built without third-party concurrency libraries — every synchronization primitive, eviction structure, and replication protocol is handcrafted and verified under concurrent load.

Attribute	Details
Stack	Java · Consistent Hashing · LRU/LFU Eviction · Multi-Node Replication
Scale	Multi-node cluster · master-slave topology · automated failover
Performance	O(1) get/set latency · intrusive doubly-linked list for O(1) eviction
Reliability	Heartbeat-based failover · monotonic epoch split-brain prevention · strict state consistency
Architecture	Thread-safe custom HashMap · LRU + LFU eviction policies · consistent hash ring
Repository

AeroCache was built as a systems correctness exercise: every component from the hash map to the replication log was written from scratch to deeply understand what production caches actually guarantee. The split-brain prevention mechanism uses monotonic epochs — each master promotion increments a global epoch, and any node receiving writes from a lower epoch rejects them silently. This ensures that in a network partition scenario, only the highest-epoch master can commit writes, maintaining linearizability across cluster membership changes.

A production-hardened code evaluation platform that securely executes untrusted C++ and Python submissions inside fully isolated Docker sibling containers with Linux privilege dropping, combined with an LLM-powered static analysis layer that delivers architectural feedback and complexity reviews instantly to the user interface.

Attribute	Details
Stack	React · Next.js · Python · FastAPI · Docker · Redis · Celery · Ollama · PostgreSQL
Security	Docker SDK sibling container isolation · fully dropped Linux capabilities · zero host exposure
Scale	Asynchronous job pipeline · Redis message broker · Celery workers · non-blocking API gateway
AI Layer	Local LLM (Qwen 2.5 Coder) via Ollama · automated static analysis · time/space complexity review
Frontend	Web-based code editor · real-time execution polling · markdown-rendered AI feedback
Repository

The security model uses Docker's SDK to spawn sibling containers — processes are never executed on the host PID namespace. Each container has all Linux capabilities dropped (--cap-drop=ALL) with a strict seccomp profile, a read-only filesystem (except a mounted /tmp), and a hard CPU + memory cgroup ceiling. The async pipeline means the FastAPI gateway never blocks on execution: submissions are enqueued to Redis, Celery workers pop and execute them in isolated containers, and results are persisted to PostgreSQL. The frontend polls a status endpoint and renders AI feedback in markdown as soon as the LLM layer completes its analysis pass.

A clinical-grade computer vision system for precise pupil and limbus boundary detection, optimized for both real-time video streams and recorded clinical footage. Achieves sub-100ms inference latency through quantized ONNX export, with mean detection error within 2mm on clinical imaging datasets.

Attribute	Details
Stack	Python · PyTorch · ONNX Runtime · OpenCV · Tkinter · Git LFS
Performance	Sub-100ms inference per frame · quantized ONNX export · mean error <2mm on clinical data
Preprocessing	Reflection removal · grayscale normalization · ring-aware preparation · ring classifier
Stability	Temporal smoothing · ROI tracking · stable detection across camera + recorded workflows
Deployment	End-to-end annotation, training, validation, and ONNX export · versioned production bundle
Repository

The system addresses a core challenge in clinical eye imaging: specular reflections from corneal surfaces corrupt pixel-level boundary detection. The multi-stage preprocessing pipeline first suppresses reflections using inpainting-guided masking, applies grayscale normalization calibrated for clinical lighting variability, and then applies ring-aware morphological preparation before segmentation. A secondary ring classifier disambiguates limbus-ring artifacts from true limbus boundaries at inference time. Temporal smoothing over a sliding window prevents frame-level jitter in live video, and ROI tracking locks the detection region between frames to avoid full-frame scanning overhead.

A real-time interpersonal harassment detection system built on YOLOv9 + SORT + LSTM, designed for edge-compatible deployment. Analyzes behavioral dynamics across sliding temporal windows and computes engineered spatial features per tracked person pair to trigger confidence-thresholded alerts.

Attribute	Details
Stack	Python · PyTorch · YOLOv9 · ONNX Runtime · FastAPI · OpenCV · SORT · LSTM
Detection	YOLOv9 person detection · SORT multi-person tracking · 30-frame sliding temporal window
Features	Normalized interpersonal distance · hand intrusion score · arm-movement sync score
Threshold	85%+ harassment confidence for high-severity alert trigger
Privacy	Face-blurring preprocessing layer · privacy-preserving by design
Deployment	Async FastAPI endpoints · edge-compatible · resource-constrained hardware targets
Repository

The behavioral model operates on pair-wise dynamics rather than individual pose estimation: for each tracked person pair within a 30-frame window, three spatial features are computed — normalized interpersonal distance (accounting for camera perspective), a hand intrusion score (detecting encroachment into a person's proximal space), and an arm-movement sync score (detecting mirroring, a behavioral indicator of coercive interaction). The ONNX-optimized LSTM classifies the sequence of these features and gates alert generation at 85% confidence. A privacy-preserving face-blurring layer is applied before any frame enters the inference pipeline, ensuring no identifiable facial data is processed or stored.