Victor Kwong

Developed an agentic audio-editing framework with improved task planning and tool scheduling, orchestrating editing operations from natural-language instructions.

Benchmarked against SmartDJ and WavCraft on under-10-second subsets of MMAE and SpeechEditBench, achieving approximately 7% improvement in evaluated benchmark performance.

Built native tracking agents in Swift (NSWorkspace + Accessibility) and C# (.NET 8 Native AOT, GetForegroundWindow + GetLastInputInfo) that write window activity to local SQLite and pause on lock, sleep and idle.

Served a Next.js static-export dashboard from each agent on a localhost API protected by a pairing code, with a timeline, sessions, categories and a monthly calendar that holds 60 fps without charting libraries.

Designed an evidence-ordered categorisation engine (user override, browser site, app name, community votes, OS metadata, window title) validated on 115 real process names and a held-out set.

Shipped per-window forgetting, timed tracking pauses and daily update notices, and designed opt-in community categories on Supabase that require 5 installs with 70% agreement.

Automated CI builds and tests for the Swift, C# and TypeScript components, attaching macOS and Windows bundles to tagged releases.

Built a rally-detection pipeline that fuses audio, motion and pose features into a gradient-boosting rally-state model, then a trajectory-aware segmenter that owns the final cut boundaries.

Combined YOLO11 shuttle detection, TrackNetV3 sequence tracking and InpaintNet gap repair into hybrid shuttle tracking that supplies boundary and score evidence.

Shipped two editions: Browser Quick analyses and exports MP4 entirely in the browser with FFmpeg WebAssembly, while Native Pro runs the full Python pipeline locally with source-quality rendering.

Built a bilingual Studio editor with frame nudging, clip cutting, undo/redo and atomic timeline saves that never overwrite human-reviewed ground truth.

Implemented a concurrent C++ B+ tree with optimistic read-latch traversal and leaf-level write locking, falling back to ancestor write latches for node splits and structural changes.

Built buffer-pool management with atomic pin counts, shared/exclusive page latches and move-only RAII guards; implemented ARC using recency/frequency lists and adaptive ghost-history tracking.

Implemented queued disk I/O on a background worker with promise/future completion, coordinating page reads, dirty-page writeback and frame replacement.

Implemented snapshot reads through MVCC undo-log reconstruction, transaction commit/abort handling, scan-predicate validation for serializable transactions and watermark-based garbage collection.

Engineered a concurrent, multithreaded caching HTTP proxy in C, combining network request forwarding with shared response caching for simultaneous client requests.

Developed a custom dynamic memory allocator using explicit free lists and boundary-tag coalescing, implementing malloc, free and realloc to manage heap blocks and reuse freed memory.

Built a Unix shell with job control, supporting foreground/background execution and process management; implemented a cache simulator to study memory-access behavior.

Completed Attack Lab exercises in x86-64 code injection and return-oriented programming (ROP), analyzing stack layout and control flow with GDB.

Built multi-source financial-data ingestion with PostgreSQL, DuckDB and Parquet storage, including validation before downstream research.

Implemented six factor families spanning risk, momentum, cointegration, crowding, entropy and formula alphas, with VectorBT parameter sweeps for candidate screening.

Implemented broker adapters for order management and position reconciliation in a paper-trading environment.

Designed a pre-deployment validation layer to separate strategy research from execution and capital-allocation decisions.

Built a Python CLI exposing composable financial-research tools for SEC filings, XBRL statements and calculations, with structured JSON and readable Markdown outputs.

Preserved source identifiers, calculation inputs and methods in tool outputs so users and AI agents can inspect the evidence behind financial analysis.

Packaged research commands with installation guides and optional OCR, table-extraction and VectorBT backtesting dependencies for different workflows.

Built a LangGraph plan-retrieve-generate workflow with semantic output checks and a bounded repair step for financial question answering.

Developed a React/TypeScript interface and FastAPI backend, integrating Supabase/PostgreSQL with Chroma for relational and vector retrieval.

Validated p95 first-token latency of 2 seconds or less and approximately 20-second agentic workflows; combined relational and vector retrieval with query latency below 200 ms.

Implemented Stripe webhooks with idempotency and retry handling, supporting 99.9% reliable real-time subscription updates and integrated payment workflows.

Built a Kafka/Flink pipeline for news ingestion, deduplication and storage, with checkpoint recovery and idempotent sinks to support failure handling.

Applied SimHash to identify similar articles across sources and automated content enrichment with SQL and JavaScript workflows.

Automated multi-tier infrastructure deployment with Terraform on AWS and Azure, managing repeatable cloud provisioning through infrastructure as code.

Implemented horizontal scaling and load balancing for web applications, with CloudWatch monitoring and automated fault-recovery strategies.