← Back home
Research and Engineering Notes
Research and engineering notes.
Notes on AI reliability, retrieval evaluation, learning systems, research method, and technical decisions.
ORCID iD · reports, preprints, and DOI-linked outputs ↗DOI-Linked Technical OutputsEarly research outputs deposited with persistent DOI identifiers. These are non-peer-reviewed technical preprints and reports.04 outputs
01Technical preprint · non-peer-reviewedFieldTrace-OCR: A Reliability-Oriented Evaluation Protocol for OCR-Based Invoice Parsing PipelinesReliability-oriented OCR invoice-parsing evaluation protocol with traceability and failure analysis.10.5281/zenodo.20669709↗02Technical report · non-peer-reviewedA Lightweight Multilingual Semantic Search Pipeline for Document Discovery and Question RoutingA modular multilingual document-discovery and routing pipeline spanning retrieval, ranking, and evaluation choices.10.5281/zenodo.20652874↗03Technical report · non-peer-reviewedTemporal Context and Momentum-Aware Reward Shaping in Reinforcement Learning: An Auditable Exploratory Study of AirRaid and MountainCarAn exploratory reinforcement-learning report focused on temporal context, reward shaping, and auditable benchmark interpretation.10.5281/zenodo.20669675↗04Technical preprint · non-peer-reviewedA Retrieval-Augmented Financial Advisory Prototype with Evidence-Gated Response GenerationA RAG prototype with retrieval, response gating, abstention boundaries, and non-product positioning.10.5281/zenodo.20652567↗
9 min read
Evaluating AI reliability beyond headline accuracy
Why production AI evaluation must cover the complete decision pipeline: ingestion, extraction, transformation, validation, review, and operational constraints.
Read note6 min read
From keyword matching to semantic retrieval
Why TF-IDF remains useful, where dense embeddings help, and how retrieval quality should be evaluated before generation is added.
Read note5 min read
What sparse rewards teach us about system design
Reward shaping is not a shortcut. It is an interface-design problem between an objective and a learning algorithm.
Read note