01
Map the system
Understand
Identify inputs, assumptions, constraints, failure modes, and the path around the model before claiming progress.
Evidence surface
OCR reliability, governance analysis, field technology support
Sydney, NSW, AustraliaReviewed · 21 Jun 2026
Current · MIT (AI) final assessments complete · formal documents pending
Applied AI Researcher and AI Systems Engineer
Shaurav Khadka
Systems lens
Understanding, predicting, and controlling complex systems under uncertainty.
I use AI, evaluation, and systems engineering to map messy inputs, model changing behaviour, and convert evidence into inspectable workflows. The same operating lens connects my applied systems work with my preparation in scientific machine learning and quantum-device characterisation.
Operating Thesis
Understanding, predicting, and controlling complex systems under uncertainty.
The portfolio is not a pile of unrelated projects. It is one trajectory: take uncertain inputs, make behaviour measurable, and turn results into systems that can be inspected, adapted, and improved.
02
Measure behaviour
Predict
Use baselines, retrieval evaluation, temporal modelling, and reproducible benchmarks to test whether the system is actually improving.
Evidence surface
Semantic retrieval, temporal graphs, RL benchmarks
03
Close the loop
Control
Translate evidence into thresholds, adaptation loops, review boundaries, and workflows that remain useful outside the demo.
Evidence surface
Sim2Real adaptation, document workflows, deployment-aware evaluation
Measured Highlights
Start with what changed.
Three benchmarks across deployment adaptation, reinforcement learning, and retrieval.
Deployment-specific Sim2Real adaptation
2.38% → 95.24%
Baseline2.38%
Adapted95.24%
Robot-image accuracy after deployment-specific Sim2Real adaptation
The model looked strong on curated data and degraded sharply on robot-camera images. The recovery came from treating domain shift as a deployment problem, not a footnote.
Why it matters: the adaptation restored useful robot-camera performance under changed lighting, viewpoint, scale, and background conditions.
Robot-camera predictions · published team-level result
Evaluation context
- Baseline
- 2.38% before deployment-specific adaptation.
- Measured
- 95.24% robot-image accuracy after targeted collection, augmentation, and fine-tuning.
- Conditions
- Robot-camera inputs with lighting, viewpoint, scale, and background differences.
- Attribution
- Collaborative team-level result with exported notebook figures.
Supporting benchmark 02
300 → 1,925
AirRaid PPO mean reward after temporal observation changes
Observation design materially changed what the policy could learn. Frame skipping and frame stacking improved the benchmark result without pretending algorithm choice was the only lever.
Why it matters: the result shows that observation design can materially change what a policy learns before the algorithm itself is replaced.
Supporting benchmark 03
P@5 = 0.68 · R@5 = 0.68
RedditPulse semantic retrieval quality
The retrieval layer was measured before generation was treated as useful. That matters because grounded insight quality depends on which sources the system surfaces first.
Why it matters: downstream summaries are only as useful as the source material retrieved before generation begins.
Working Method
Understand. Predict. Control. Reproduce.
The operating loop behind the portfolio: turn uncertainty into evidence, then turn evidence into a system decision.
- 01
Frame
Define the system boundary, decision path, assumptions, constraints, and failure cost before choosing a model.
- 02
Measure
Build baselines, evaluation splits, success conditions, and instrumentation that make behaviour observable.
- 03
Stress-test
Probe edge cases, domain shift, confidence failures, transformation errors, and human-review boundaries.
- 04
Control
Adjust thresholds, representations, workflows, or learning setup only where the evidence justifies intervention.
- 05
Reproduce
Preserve the data boundary, method, result, limitation, and decision record so the claim can be inspected later.
Applied Systems
Selected applied systems.
Production reliability, temporal learning, retrieval, evaluation, and responsible-AI work.
Prior industry workflowTRUUTH · Former AI/ML R&D Internship
Production AI Reliability and Document Intelligence
Problem: Document intelligence can fail long before or after OCR. Real reliability depends on the complete path from ingestion to extraction, transformation, validation, and review.
Contribution: Built repeatable evaluation workflows across OCR configurations, mappings, confidence scores, error codes, and reruns while preserving traceability and review boundaries.
PythonpandasAWS S3boto3Azure Document IntelligenceJSON
Inspect case study 01
OCR
02
Map
03
Validate
04
Trace
Shared here: sanitised workflow record. Confidential operational data and internal metrics are excluded.
Academic and industry referees are available on request for selected roles and research collaborations.
Supporting builds
Five supporting systems across research, prototypes, and evaluation.
Each card states the system, category, and route for deeper inspection.
Temporal Graph Learning
Temporal graph-learning research build
Temporal GNN for Blockchain Fraud Detection
Fraud is relational and time-dependent. Static tabular features can miss how transactions evolve across a network.
t0 → t1 → t2
Inspect case studyGenerative AI · Conversational Systems
Scoped conversational-AI prototype
LLM-Based Financial Assistant Prototype
Conversational assistants can produce fluent but poorly scoped responses. This prototype explores structured prompting, model comparison, synthetic profiles, and explicit safety boundaries.
profile → prompt → compare → respond
Inspect case studyNLP · Information Retrieval
Modular retrieval research toolkit
R01
R02
R03
Semantic Search and Information Retrieval Engine
Keyword matching is transparent but limited when meaning varies across phrasing. The system needed a modular comparison path from classical retrieval to dense semantic search.
clean → encode → rank → evaluate
Inspect case studyMachine Learning · Data Science
Reusable experimental evaluation pipeline
Benchmark 1
Benchmark 2
Benchmark 3
Benchmark 4
Benchmark 5
Machine-Learning Evaluation and Data-Science Pipeline
A model result is only useful when the path from raw data to evaluation is reproducible, comparable, and explicit about failure cases.
data → features → compare → inspect
Inspect case studyResponsible AI · Governance
Responsible-AI research and analysis portfolio
Risk
Signals
Oversight
Improve
Responsible AI, Governance and Human-Centred Analysis
AI systems can be technically capable and still fail users, organisations, or communities when accountability, transparency, risk, and human oversight are treated as afterthoughts.
risk → explain → govern → improve
Inspect case studyProject index
Browse all project routes.
A compact index of results, methods, and prototypes.
09 routes
Project index
Browse all project routes.
A compact index of results, methods, and prototypes.
Filter project routes
Use the filters to browse the complete project catalogue.
Filter index
Showing all project routes
Category
Published results
Exported figures, media, and measured evaluation outputs.
Category
Methods
Sanitised workflows and documented research methods.
Category
Exploratory prototypes
Prototype architectures and exploratory implementations.
Research Profile
Research Directions
My research direction follows the same operating thesis: understand complex systems, predict behaviour under uncertainty, and control outcomes through careful evaluation.
The longer-term direction that most motivates me is AI-assisted quantum-device characterisation. I am building the mathematical and physical foundations carefully: study the theory, reproduce small experiments, test implementations, and make stronger claims only when the work earns them.
Discuss a research collaborationEmerging research trajectory
AI-Assisted Quantum Device Characterisation
A longer-term interdisciplinary direction focused on noisy quantum devices as complex systems under uncertainty. The bridge is open quantum systems, temporal reasoning, scientific machine learning, and careful experimental validation.
Questions I am building toward
- Where can temporal AI methods assist characterisation without obscuring the underlying physics?
- Which mathematical, simulation, and experimental foundations must be reproduced before stronger claims are justified?
Current research priority
Trustworthy AI Systems and Production Reliability
Evaluation methods for AI pipelines where traceability, robustness, auditability, confidence handling, latency, and cost matter alongside model accuracy. My prior document-intelligence internship work treated the full decision pipeline, not an isolated model, as the unit of analysis.
Questions I want to pursue
- How should reliability be measured across the full decision pipeline?
- How can failure analysis distinguish model, data, and system faults?
- How should confidence, review boundaries, and operational constraints shape deployment decisions?
Active research area
Temporal Learning for Dynamic and Relational Data
Learning systems for data that evolves over time: temporal graphs, sequential signals, changing relationships, and non-static risk patterns.
Questions I want to pursue
- When does temporal modelling materially outperform static baselines?
- How should time-dependent behaviour be evaluated and explained?
Active research area
Sim2Real Perception and Autonomous Systems
Robust perception under deployment shift, confidence-aware decisions, and vision-to-action systems that must behave safely outside curated datasets.
Questions I want to pursue
- How can deployment adaptation be designed in from the beginning?
- How should confidence thresholds shape downstream actions?
Research mapApplied foundations and next stepsThree foundation groupsOpen map
The bridge is concrete: systems work in reliability, temporal modelling, retrieval, deployment adaptation, and focused scientific-ML preparation.
Reliability and evaluation
Applied methods supported by inspectable systems work and explicit benchmark results.
- Traceability
- Confidence analysis
- Error analysis
- Benchmark design
- Operational constraints
- Auditability
Learning under change
Built foundations for dynamic data, deployment shift, and reproducible experimentation.
- Semantic retrieval
- Temporal modelling
- Graph learning
- Sim2Real adaptation
- Uncertainty analysis
- Reproducible evaluation
Scientific-ML preparation
Foundational preparation for physically informed AI research through reproducible study and small implementations.
- Scientific machine learning
- Dynamic systems
- Computational modelling
- Hybrid-QML exploration
- Quantum-device characterisation · emerging
Experience
Research and Technical Experience
Applied research, production-oriented AI R&D, technical leadership, and software engineering.
TRUUTH
AI/ML Research and Development Intern
Feb 2026 to Jun 2026
Sydney, NSW, Australia · Hybrid
Production-oriented document intelligence, fraud-detection evaluation, and AI reliability analysis. Built repeatable OCR-evaluation workflows across layouts, configuration choices, confidence scores, field mappings, and error codes while documenting traceability, reproducibility, validation dependencies, latency, and cost considerations.
- Document Intelligence
- OCR Evaluation
- AWS S3
- Azure Document Intelligence
- Reliability
Picpoint Nepal Pvt. Ltd.
Chief Technology Officer
Jun 2021 to Jun 2024
Kathmandu, Nepal · Hybrid
Technical leadership across operational systems, digital workflows, and data-informed decision support. Led the technical roadmap and maintained systems supporting remote workflows, business coordination, web operations, and market-intelligence tooling.
- Technical Leadership
- Operations Systems
- Data Workflows
- Web Systems
Thakur International
Jr. Full Stack Developer
Jun 2019 to May 2020
Kathmandu, Nepal · On-site
Application development, API integration, debugging, and backend-data quality within an agile engineering team. Implemented and maintained web and mobile components while improving maintainability through structured debugging, refactoring, and performance tuning.
- PHP
- Python
- JavaScript
- REST APIs
- Debugging
Additional experienceSupporting work completed during postgraduate study
Ingleburn Convenience Store
Operations and Digital Support Assistant · Part-time
Oct 2024 to Jun 2026
Supported transaction and inventory accuracy, POS troubleshooting, basic network and hardware issues, digital administration, and customer-facing operations during postgraduate study in Australia.
Foundation
Education and Continuing Study
Formal academic foundation, degree-completion update, and selected non-credit certificates that support current research preparation.
Education
Macquarie University
Master of Information Technology · Artificial Intelligence
2024 to 2026 · Sydney, NSW, Australia
Coursework and final examinations completed; official completion letter and transcript pending. Expected overall result: Distinction, pending final university release. Relevant work: NLP and LLM systems, graph machine learning, advanced computer vision and action, reinforcement learning, AI governance, and an industry AI/ML R&D internship.
Education
London Metropolitan University · Islington College
BSc Computer Science · First Class Honours
2017 to 2021 · Kathmandu, Nepal
Ranked among the top 10 students in the cohort. Built recommendation, trip-planning, and database-backed systems across Python, PHP/MySQL, Oracle, C#, Java, and GUI development.
Selected Continuing Study
Non-credit online certificates added as supporting preparation, not as university degree credit.
Coursera · Jun 2026
Understanding Modern Physics II: Quantum Mechanics and Atoms
The Hong Kong University of Science and Technology
Foundational quantum mechanics and atomic-physics study supporting research preparation in open quantum systems and quantum-device characterisation.
Verify NGRUYQGSMAMJCoursera · Jun 2026
Quantum Mechanics
University of Colorado Boulder
Foundational quantum-mechanics certificate used as targeted preparation for scientific-machine-learning and quantum-characterisation work.
Verify 861MS6P64X8ICommunity Impact
Community Technology and Leadership
Field technology support and selected leadership programs connected to sustainability, peer guidance, design thinking, and cross-cultural collaboration.
Long-term field technology
Technical Volunteer and Systems Support
Since
2015
Swogun EnergyNepal and remote support · Periodic volunteer contribution
Supported field deployment, testing, and troubleshooting of small-scale solar-power and IT systems in remote and off-grid settings in Nepal. Continues to provide occasional remote technical and digital support while based abroad.
- Solar Systems
- Field Support
- IT Troubleshooting
- Remote Communities
Why it belongs here
This is not a decorative add-on. It shows practical technology work under constrained, field-facing conditions: power systems, troubleshooting, remote support, and community deployment.
About
Complex Systems, Tested Before They Are Trusted
I am an applied AI researcher and AI systems engineer focused on understanding, predicting, and controlling complex systems under uncertainty.
My work spans document intelligence, semantic retrieval, temporal graph learning, computer vision, reinforcement learning, and production-oriented evaluation. I care about the full path around a model: inputs, representations, benchmark design, failure analysis, review boundaries, and the workflow that eventually reaches users.
Longer-term interests include scientific machine learning and quantum-device characterisation. I approach them through careful study, small reproducible experiments, and stronger claims only when the evidence earns them.
Research stance
I do not trust a result I cannot inspect.
I do not treat a benchmark as evidence until it survives failure cases.
I do not publish a claim I cannot reproduce.
The rest is disciplined research.
Inspect
Expose the assumptions, inputs, and boundaries behind each result.
Stress-test
Look for failure cases before treating a benchmark as evidence.
Reproduce
Keep claims tied to work that can be checked, rerun, and improved.
Research Writing
Research Notes and Engineering Decisions
Four on-site notes on research questions, evaluation choices, and engineering decisions, with four DOI-linked technical outputs below.
04 notesORCID iD ↗
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↗
Independent Publishing
Books & Independent Publishing
Independent authorship, illustration, and editorial credits presented as a compact publishing record.
Featured authored publication
The Digital Equilibrium
Navigating Technological Advancement for Optimal Well-Being
An independent authored work exploring how technological progress can be balanced with human well-being and intentional living.
A small publishing trail spanning technology, well-being, and selected creative collaboration.
Collaborative editions6 selected creative credits
Illustrated and editorial work
Selected illustration and editorial credits across children’s stories and reflective writing.
Browse titles
Illustrator · Editor
Children Stories
Bal Katha
Illustrator · Creative contributor
Joyful Stories
Joyful Stories
Illustrator · Editor
Joyful Stories
Mazzako Katha
Illustrator · Creative contributor
Joyful Stories
Mazzako Katha · Alternate edition
Illustrator · Editor
Words of Wisdom
Amritvani
Illustrator · Creative contributor
2 in 1 Joyful, Children Stories
Combined children’s-story edition
Contact
Choose the conversation worth continuing.
Choose a research or role-focused conversation, or continue exploring the portfolio.
Research
Research collaboration
Exploring dependable AI systems, deployment-aware evaluation, or scientific-ML directions? Start with the notes or open a research conversation.
Roles
Applied AI research and engineering roles
Hiring for applied AI research, ML systems, or reliability-focused engineering? Review the résumé, inspect the systems portfolio, and start a focused conversation.
Continue exploring
Explore more of the portfolio before reaching out.
Current status
Sydney-based and open to selected applied AI systems roles and research collaborations while formal Macquarie completion documentation is pending.
Response target
I aim to reply within one to two business days.
Referees available
Academic and industry referees are available on request for selected roles and research collaborations.