projects.txt

Projects

Things I've built, on my own time and on the job.

Personal

CruxCam

Source Live

Engineered an inference pipeline using MediaPipe BlazePose to track 33 3D landmarks per frame and classify climbing positions by joint angle in metric world space
Implemented a biomechanics-based center-of-mass estimator using anthropometric segment weights, EMA temporal smoothing, and torso-scale pixel projection
Architected an async backend with FastAPI and Redis, dispatching video processing jobs to a ThreadPoolExecutor
Optimized video throughput with a multithreaded reader–inference–writer pipeline overlapping disk I/O with model inference, re-encoding output to H.264 via ffmpeg
Built an interactive 3D skeleton viewer with React Three Fiber synced to video playback; extended to a side-by-side compare mode with shared play/pause controls, independent per-clip trim and efficiency scores, and parallel job polling
Designed a transport-agnostic core package keeping inference logic decoupled from API and worker contexts

PythonComputer VisionMediaPipeFastAPIRedisReact Three Fiber

BEVFusion

Source

Implemented a camera–LiDAR fusion pipeline for 3D object detection in PyTorch based on the BEVFusion (MIT CSAIL, 2022) architecture, evaluated on the nuScenes benchmark
Built a Lift-Splat-Shoot camera encoder using EfficientNet and a learned depth distribution to project 6 synchronized camera views into a shared Bird's Eye View grid
Engineered a PointPillars LiDAR encoder, pillarizing raw point clouds and applying a shared PointNet MLP to produce a matching BEV feature map
Designed a fusion encoder and SSD-style detection head to concatenate camera and LiDAR BEV features and output 3D bounding boxes, class labels, and headings
Developed a full training pipeline with a custom nuScenes dataloader, anchor matching, and focal loss, plus inference scripts for NMS and height-colored BEV visualization
Targeted ~67% mAP / 71% NDS on nuScenes val by fusing modalities, versus ~35–50% mAP for camera-only or LiDAR-only baselines
Authored in-depth technical documentation explaining the LSS depth-lifting mechanism and why it learns accurate depth without direct supervision

Computer VisionSensor FusionPyTorchLiDARnuScenes

FSRCNN Super-Resolution

Source Demo

Implemented a PyTorch-based FSRCNN super-resolution model and trained it end-to-end on custom image datasets
Built reusable data loading and preprocessing pipelines to support efficient training and evaluation
Achieved PSNR within ~2 dB of the paper's reported benchmark results, validating implementation correctness
Deployed the model via a Flask service and containerized the system with Docker for reproducibility

Computer VisionPyTorchFlaskDocker

SpaceX CRS-10 Telemetry Analysis

Source

Investigated three unsupervised anomaly detection methods (Z-score, IQR, Rolling Z-score) on high-frequency SpaceX CRS-10 rocket telemetry, using 5 known flight events as ground truth
Built a cluster-based evaluation framework to group detected anomalies and score each method on precision, recall, and F1, enabling direct comparison across techniques
Discovered through grid search that the optimal Z-score threshold (0.7) was far below the conventional 2.0–2.5 range, revealing the signal's compressed variance as a key driver
Concluded that locally-adaptive methods consistently outperform global statistics on non-stationary signals, a finding with direct implications for real-world sensor data pipelines

PythonTime-SeriesAnomaly Detection

Work

Asset Management Platform

Generate Capital

Built a full asset management system that ingests behind- and front-of-the-meter power data into Snowflake and carries it straight through to reporting, managing 300+ projects
Gives the team insight into how ESA contracts and assets are performing, as an internal tool hosted on Azure with a JavaScript front end
Includes performance forecasting and anomaly detection, supporting proactive maintenance and capital planning decisions

SnowflakeAzureJavaScriptForecastingAnomaly Detection

Unplanned Maintenance Analysis

Generate Capital

Designed the analysis pipeline, joining a maintenance work-order tracker with Snowflake revenue and site-mapping data to measure how unplanned maintenance affects site-level profitability
Modeled each site's failures as a Poisson process to turn failure history into a forward-looking expected value at risk
Built a size-independent reliability metric (work orders per generator) and paired it with site value to separate genuinely unreliable sites from merely large ones
Designed a rule-based triage framework combining financial, recurrence, and risk signals into review, protect, and monitor tiers
Tested seasonal cost assumptions with Kruskal-Wallis and Mann-Whitney U tests, traced repeat failures to their downstream escalation, and documented assumptions, caveats, and data gaps in a written report

PythonSQLSnowflakeRisk ModelingStatistics

Billing Data Review Workflow

Generate Capital

In progress

Adding a feature to an internal billing tool that catches faulty performance data and routes it for manual review
Building the review UI and designing the system underneath that routes flagged data through the workflow
Setting security permissions for the feature and working across teams to meet stakeholder needs

System DesignUI DevelopmentAccess ControlData Quality

DSCR Data Pipeline in dbt

Generate Capital

Moved the Debt Service Coverage Ratio (DSCR) logic out of a Power BI report and down to the dbt level, automating a calculation stakeholders previously did by hand
Used SQL to merge multiple data sources and compute DSCR in the warehouse rather than in Power BI
Pushing the logic upstream improved efficiency and accuracy and made the resulting tables reusable by other areas of the company
Left a leaner, more efficient dashboard with fewer opportunities to fail, since the merging and calculation logic no longer lives in Power BI

dbtSQLData EngineeringPower BI

AI Report Summarization System

BPM LLP

NLP pipeline integrating Azure Cognitive Services and GPT-4 to summarize 100+ page transfer pricing reports, cutting analyst preparation time by ~98% (4 hours → ~5 minutes)

PythonLLMNLPAzure

Invoice Reconciliation Automation

BPM LLP

Applied fuzzy matching to reconcile invoice line items against project estimates and generate automated PDF outputs, reducing processing time from ~5 hours to under 1 minute

PythonFuzzy MatchingAutomation

Renewable Energy Transfer Pricing Model

BPM LLP

Built a simulation model to optimize intercompany renewable energy transfer pricing, maximizing profit through carbon credit monetization

PythonSimulation

Personnel Operations Dashboard

BPM LLP

Power BI dashboard built for a professional association client to track personnel metrics, headcount, and operational KPIs for leadership decision-making

Power BISQLPython