Wael

Solution for the 2026 IEEE Big Data Cup TrafficFlowBench on ten synthetic freeway corridors: reconstruct blanked traffic states, forecast queues 30 minutes ahead, respect traffic physics and recover origin-destination demand.

• Strictly causal queue forecasting: every input is stamped at or before the forecast origin T. A deletion test blanks every cell after T, rebuilds inputs and predictions, and they come out bit-identical; a negative control shows the test catches leaks.
• State reconstruction with an XGBoost residual model plus a 14-network 2-D convolutional imputer ensemble and a shock post-correction.
• Queue forecasting with a space-time CNN for ongoing queues and a LightGBM + CatBoost onset blend with a "month-pure" memory of earlier days.
• Origin-destination demand recovered with non-negative ridge regression; physics treated as an identity of the state answer, not a separate target.
• Spotted a regime shift between evaluation windows that the public leaderboard could not show, and documented it in a technical report.

Stack: Python · PyTorch · XGBoost · LightGBM · CatBoost · scikit-learn · pandas · NumPy

Repository

A relay of three LLM agents that turns buggy Python into clean, tested code. An Auditor finds issues with Pylint and an LLM, a Fixer rewrites the code with Gemini, and a Judge generates and runs tests. Failing tests send the error back to the Fixer in a self-healing loop.

• Orchestrated as a LangGraph state graph with conditional edges (plus a legacy relay orchestrator).
• Listener (Auditor): Pylint static analysis plus LLM issue detection. Corrector (Fixer): Gemini-generated fixes. Validator (Judge): LLM-generated pytest suites.
• Sandboxed file tools with path validation, and JSON telemetry of every prompt and response with an HTML dashboard.

Stack: Python · LangGraph · LangChain · Gemini · Pylint · Pytest

Repository

The challenge's scored road layer drops service and pedestrian roads ("escape-route framing"). In the network it defines, 2.88 million residents (7.1%) live in 13,775 pockets whose only connection is a single road link, and the coverage score never counts ways out.

• For 17.5% of pocket residents in the highest wildfire-hazard quartile, a second way out exists in Overture Maps, but only in the road classes the filter drops.
• Concentrated where the hazard is highest, and checked against OpenStreetMap and the FHWA public-road inventory.
• Shows who carries it (care homes, residents 65+, mobile-home parks) and what an egress component would change in the scorecard.
• Delivered an action list of 5,095 high-hazard single-exit pockets; a verification script recomputes all 107 reported numbers.

Stack: Python · pandas · DuckDB · Shapely · pyproj · Overture Maps · OpenStreetMap

Repository

Foundational algorithms implemented from the ground up in Python notebooks, to understand the mechanics behind the libraries: linear and logistic regression, decision trees and random forests.

• Linear regression trained with gradient descent on the cost function.
• Logistic regression with a sigmoid and cross-entropy loss.
• Decision trees with feature-based splits, and random forests over random data subsets.

Stack: Python · NumPy · Jupyter

Repository

A small, dependency-free tool and live dashboard for participants. It reads the public leaderboard and shows where headline points are still unclaimed per metric, estimates where a submission stands, and tracks how the field moves.

• Reproduces the published headline formula from the task weights and prints the check on every report.
• A "where do I stand" calculator runs fully in the browser and sends nothing anywhere.
• A daily GitHub Action stores an anonymised snapshot (team names and affiliations are dropped, enforced by a test).

Stack: Python (stdlib) · GitHub Actions · GitHub Pages · HTML

Repository · Live dashboard

A lightweight fine-tuning workflow for Qwen2.5-3B-Instruct: clean the UltraChat conversations, train a 4-bit QLoRA adapter on a single GPU, publish checkpoints to the Hugging Face Hub and compare base vs fine-tuned generations.

• Data cleaning: turn and length filters, strict user/assistant alternation, de-duplication and normalisation.
• 4-bit NF4 quantisation (BitsAndBytes), LoRA rank 16 / alpha 32, paged 8-bit AdamW.
• Checkpoints pushed to the Hugging Face Hub; inference script compares the base model and the adapter.

Stack: PyTorch · Transformers · PEFT · BitsAndBytes · Hugging Face Hub

Repository

A Next.js + TypeScript web app where carriers book terminal appointments for their containers, and operators and admins manage bookings, terminals, carriers, notifications, reports and an audit log.

• Role-based portals (admin, operator, carrier) plus chat and notifications.
• A centralised API layer with CSRF-token handling and HTTP-only cookie sessions.

Stack: TypeScript · Next.js · React · Tailwind CSS · Radix UI · socket.io

Repository