M. Hassan Shahzad

• Built automation pipelines collecting shipment tracking data from 10+ carrier portals, extracting delivery status and logistics metadata for daily cargo monitoring
• Developed an AI-driven document processing pipeline using the Gemini API to convert heterogeneous, unstructured document formats into standardised structured datasets
• Engineered a cargo booking automation system submitting bookings across multiple carrier portals via APIs and scripted workflows, centralising them into a single internal system

• Contributed to 3 open-source codebases including Gemini CLI, submitting 8 pull requests with 2 merged
• Performed code review, bug triage, and feature implementation in collaboration with upstream maintainers
• Gained experience navigating large unfamiliar codebases and reproducing reported issues

Habib University at SemEval-2026 Task 3: A Pipeline Approach for Dimensional Aspect-Based Sentiment Analysis

Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026), ACL, pp. 3449–3459 · 2026

Second author. Four-stage multilingual transformer pipeline (extraction, aspect–opinion pairing, category classification, valence–arousal regression) across six languages and four domains. Ranked 2nd internationally in Tatar and 6th in Russian; best-system RMSE 0.5333 (Subtask 1), cF1 0.5492 (Subtask 2). Designed regression stabilisation strategies (label rescaling, Gaussian label noise, CCC loss, Savitzky–Golay smoothing) cutting dev RMSE from 0.9572 to 0.9366, and ran augmentation ablations that uncovered a cross-lingual regularisation effect from external Chinese data.

HU_LLM_Fin @ FinMMEval 2026 Task 2: A Structure-Aware Hybrid RAG Pipeline utilizing Mixture-of-Experts

CLEF 2026 Working Notes, CEUR Workshop Proceedings (accepted, in press) · 2026

Joint first author. Cross-lingual QA over SEC financial tables and multilingual news in five languages: co-designed a row-aware chunking algorithm preserving table structure, two-stage partitioned retrieval fusing sparse + dense signals with a cross-encoder reranker, and a Mixture-of-Experts generation layer with scratchpad reasoning. Placed 10th globally on the official blind test set (ROUGE-1 0.3958 Easy / 0.3437 Expert).

Honors & Awards

• 2nd place (Tatar), 6th place (Russian) — SemEval-2026 Task 3: Dimensional Sentiment Regression (2026)
• 10th place globally — CLEF-2026 FinMMEval Lab, Task 2 (PolyFiQA) official blind test set (2026)
• 5th place (Tatar), 6th place (Russian and Ukrainian) — SemEval-2026 Task 3: Triplet Extraction (2026)

A locally deployed LLM assistant with tool-calling for external APIs and function execution, avoiding third-party inference dependencies. Multi-agent modular pipeline (query refinement → tool invocation → output formatting) deployed into the GAMA Suite platform serving 1000+ customers.

An adversarial search agent for the Sequence board game with multiple difficulty levels, playable in real time online. Minimax with alpha–beta pruning and custom heuristic evaluation over board states; real-time multiplayer full-stack app with a WebSocket-driven Next.js frontend.

Manual tracking and booking of cargo across many shipping-carrier portals. Automation pipelines collecting shipment tracking data from 10+ carrier portals, an AI document-processing pipeline standardising unstructured documents, and a cargo booking system centralising multi-carrier bookings.

A portfolio that proves engineering skill instead of listing it — built as an AI chat app. Streaming AI twin grounded in his résumé, with a zero-cost offline fallback engine, rate limiting, and prompt-injection handling.

B.S. in Computer Science

Relevant coursework: Introduction to Large Language Models (CS/CE 335/466), Machine Learning, Artificial Intelligence, Data Structures & Algorithms, Linear Algebra, Probability & Statistics, Database Systems.