Elijah Ng

• Designed and deployed a Prefect-based ETL pipeline consolidating 50+ legal/real estate data sources into a unified data lake; processed 3k records/day with idempotent and checkpointed runs, growing qualified leads by 10x YoY.
• Implemented OpenTelemetry-based observability with LLM-enriched diagnostics to expedite pipeline debugging
• Ship an LLM-based information retrieval and document understanding pipeline to extract validated, structured data from unstructured filings to automate downstream lead qualification
• Built a context-aware query-variant generator and results filter for skip-tracing. Improved contact hit-rate by 20x while capping cost to <$1/lead

• Owned app/frontend design and implementation; built on-chain analytics tooling for the airdrop platform

• Processed 1.3TB of raw telemetry data via Dusk to engineer features for a physics-informed maritime CO2 emissions model. Combined geospatial data with parsed vessel engineering specs to achieve explainable estimates

• Improved YOLOX-tiny model mAP from 0.60 to 0.85 through data augmentation (SimCLR), LR scheduling, implementing a HPO pipeline, and optimizer implementation tweaks
• Developed quantisation + compilation pipeline for YOLOX ONNX graphs to RKNN (Rockchip NPU)
• Initiated and led regional team on semi-supervised image labeling and dataset curation via latent space analysis (using fine-tuned GAN embeddings), saving hundreds of team hours per project

• Implemented custom Triton kernel for NF4 dequantisation, achieved 25% speedup over Unsloth baseline on T4
• Enabled QLoRA fine-tuning with FSDP2 and torch.compile with no graph breaks
• Implemented a memory-efficient backprop (inspired by cut-cross-entropy) that is compatible with GRPO

• Fine-tuned Qwen3-Embedding-4B via 4-bit QLoRA (MultipleNegativesRankingLoss) on a generated contrastive dataset spanning 700+ mechanics; deployed via llama.cpp in the embedding pipeline of the semantic search app
• Built a curated challenge-set benchmark with hard negatives and nDCG@k/MRR reports to guide iteration

• Reduced peak VRAM by 50% for Stable Diffusion fine-tuning through quantization and attention chunking. Turned client profitable by fitting training on 3080 instances from 3090s

• Doubled down on analysis over visuals: shipped a functional Jupyter Notebook with aggregated analytics for 1inch Fusion resolver on-chain activity, execution profits, and gas spend

• Curated a dataset of 2,000+ tokens from 2013–2023 across venues; developed backtests for systematic index fund strategies with modelled and simulated execution conditions (slippage, liquidity, etc.)

Open Source

• Contributed bug fixes to PyTorch Lightning and bt (backtrader)
• Classic SGD: Reverted PyTorch SGD to original Sutskever formula for separable LR/momentum behavior
• Dotfiles: collection of configs and scripts for AI-tools (OpenCode, Claude Code), NixOS, productivity tracking, etc.