Résumé

Problem: Most AI resume scorers behave like black boxes, producing inconsistent scores without showing which resume evidence supports each decision.

System: Modeled resume-JD matching as a structured reasoning pipeline across skills, experience, tools, role alignment, mandatory requirements, optional requirements, and evidence-backed scoring.

Outcome: Built a deterministic evaluation approach where LLMs support interpretation, but structured rubrics, retrieved evidence, and traceable scoring control the final judgment.

Problem: Long-form video is difficult for AI systems to reason over because raw video lacks explicit scene boundaries, temporal structure, and reusable semantic units.

System: Built an experimental pipeline using FFmpeg, scene detection, transcript segmentation, and local LLM reasoning to convert videos into structured, auditable media intelligence.

Outcome: Created the foundation for a reproducible video analysis pipeline where content generation is guided by explicit structure instead of uncontrolled creative prompting.

Problem: Doctor-patient conversations contain useful diagnostic signals, but those signals are often buried inside messy, informal, and uncertain natural language.

System: Combined NLP-based symptom extraction, vector search, and graph-style disease association modeling to organize conversational medical evidence into explainable structures.

Outcome: Built a controlled prototype for healthcare reasoning where the system focuses on traceable symptom evidence instead of direct black-box diagnosis.

Problem: Medical reports often contain dense tables, repeated formatting, irrelevant text, and scattered values that are difficult for users to understand digitally.

System: Designed a FastAPI-based extraction pipeline to parse medical PDFs, identify meaningful tabular or row-like values, ignore irrelevant content, and prepare structured data for downstream AI analysis.

Outcome: Established the backend foundation for a report intelligence product that can turn static PDFs into usable, explainable health data summaries.

Problem: Interview transcription systems become unreliable when audio is noisy, multilingual, streamed in real time, or captured under unstable network conditions.

System: Built a speech-to-text backend with audio windowing, silence detection, diarization, translation support, and failure-safe ingestion using FastAPI, Whisper, and Deepgram.

Outcome: Created a reliable transcription foundation for AI interview systems where spoken answers can be captured, cleaned, and evaluated with better traceability.

Problem: Short-form content creation is bottlenecked by repetitive manual editing, clip selection, and formatting decisions.

System: Designed an automated media pipeline that uses video processing, scene segmentation, and AI-assisted reasoning to identify usable moments and prepare short-form outputs.

Outcome: Built a research-adjacent automation system exploring how structured media understanding can reduce editing overhead while keeping content decisions explainable.

Problem: Rare network attacks are difficult to detect reliably because the dataset is imbalanced and simple accuracy can hide poor minority-class performance.

System: Built a multi-stage IDS pipeline using stacked sparse convolutional autoencoder representations, raw traffic features, clustering signals, and ensemble classifiers to compare learned evidence against strong classical baselines.

Outcome: Showed that raw XGBoost remained the strongest default model, while learned representations and hybrid evidence became useful under selected feature-stress conditions.

Problem: Single-modality diagnostic models can miss important disease signals because neurological disorders often appear across multiple behavioral and physiological patterns.

System: Designed a hierarchical fusion strategy that combines gait, handwriting, and spatiotemporal signals into a unified diagnostic representation.

Outcome: Demonstrated a research-grade diagnostic pipeline where multimodal fusion provides richer evidence than unimodal prediction alone.

Problem: Static education analytics often fail to capture how student behavior evolves over time before disengagement or withdrawal.

System: Applied dynamic time warping, unsupervised clustering, and manifold-style behavioral analysis on student activity trajectories from the OULAD dataset.

Outcome: Created a framework for detecting distinct engagement patterns and surfacing early-risk learning trajectories before final outcomes are visible.

Problem: AI systems are often treated as deterministic when seeds and temperature are fixed, but hardware-level numerical behavior can still introduce output drift.

System: Built experiments comparing CPU and Apple MPS execution paths to study floating-point non-determinism, neural activation drift, and parallel accumulation differences.

Outcome: Produced a practical foundation for auditable AI design by showing why software-level determinism must account for hardware-level entropy.