# Vicky Feliren

> Applied Scientist working on AI safety for trustworthy, multimodal, and multilingual systems

Vicky Feliren is an applied scientist specializing in AI safety and the development of trustworthy multimodal and multilingual systems. He conducts applied research that bridges vision–language models, uncertainty quantification, and interpretability, and has published in venues including IEEE, ACL, and Remote Sensing of Environment. Vicky holds a Master of Data Science and a Bachelor of Computer Science from Monash University, and contributes to initiatives around responsible AI and regionally relevant multimodal research. He has been recognized with awards such as the Global South AI Safety Hackathon (Asia Pacific Regional Winner, 2026) and the Microsoft Azure Virtual Hackathon APAC (Regional Champion, 2020). Vicky is available for research collaborations, speaking engagements, and consulting on trustworthy AI.

## Pages

- [Home](https://vickyfeliren.com/): Hero, research identity and entry point
- [About](https://vickyfeliren.com/about/): Bio, research agenda, expertise, work history, and awards
- [Research](https://vickyfeliren.com/research/): Publications, IEEE Q1, ACL 2025, Remote Sensing of Environment Q1, AQUA Q2, arXiv preprints
- [Use Cases](https://vickyfeliren.com/usecases/): Applied ML project write-ups with full SCR (Situation–Complication–Resolution) narratives across multimodal AI, multilingual and cultural NLP, earth observation, production fintech, and hackathon projects
- [Writings](https://vickyfeliren.com/writings/): Research notes (short paper distillations with own take), press features, and Medium essays on AI, philosophy, and life. Defaults to the Research filter.
- [Work With Me](https://vickyfeliren.com/contact/): Available for applied scientist roles, research collaboration, speaking, and mentorship

## Use Cases (individual pages)

### Multimodal AI · Earth Observation
- [Flood Segmentation, ProCANet](https://vickyfeliren.com/usecases/procaenet-flood-segmentation/): First-author IEEE GRSL 2025. Progressive cross-attention network for multispectral satellite flood segmentation. IoU 0.815 on Sen1Floods11; zero-shot transfer to PlanetScope across 6,112 km².
- [Mining Footprint Detection](https://vickyfeliren.com/usecases/mining-footprint-segmentation/): Remote Sensing of Environment Q1 (IF 11.4) 2025. Multi-modal deep learning fusing Sentinel-2, SAR, and DEM for continental-scale mining footprint delineation. First application of a geospatial foundation model (Prithvi) to this task.
- [Aquaculture Pond Detection](https://vickyfeliren.com/usecases/aquaculture-pond-detection/): Filed patent IDS000010594 (June 2025). Multi-temporal Sentinel-2 pipeline for illegal aquaculture pond expansion detection across Indonesia and Vietnam coastlines.
- [Flood Policy Evaluation](https://vickyfeliren.com/usecases/flood-urban-resilience/): AQUA Q2 2025. Mixed-methods evaluation of retention pond effectiveness in South Bandung, bridging ProCANet satellite segmentation with evidence-based urban policy.

### AI Safety & Reliability
- [Multilingual VLMs Under Cross-Modal Conflict](https://vickyfeliren.com/usecases/multilingual-vlm-crossmodal-conflict/): Apart Research Global South AI Safety Hackathon 2026 (Co-First Author). Multilingual (English, Hindi, Telugu) cross-modal conflict benchmark for 9 open VLMs with a paired perception-control override-gap design. The conflict signal stays linearly decodable at 0.92 in Telugu; cross-lingual contrastive steering transfers an English-fit direction that drives text-override to zero for abstention-amenable models.
- [Hidden-State Detection of In-Context Goal Hijacking](https://vickyfeliren.com/heron/): Heron AI Security Research Fellowship, work-test prototype (2026). Read-only linear probe on Qwen2.5-Instruct (0.5B-7B) hidden states detects in-context goal-hijack attempts with deconfounded AUC 0.998 and a split-conformal false-positive guarantee (0.029 vs. target 0.05). A benign-prefix confound control caught a naive detector flagging 100% of harmless prefixed prompts before it shipped, and an input-text baseline bounds what hidden states add here (held-out-family TPR 0.988 vs 0.880 for the text-only detector). Structured summary: https://vickyfeliren.com/usecases/heron-hijack-self-probe/

### Cultural & Multilingual AI
- [SEA-VL Benchmark](https://vickyfeliren.com/usecases/sea-vl-benchmark/): ACL 2025 (Main Conference). Co-led 100+ annotator crowdsourcing across 11 countries. First rigorous culturally-grounded VLM benchmark for Southeast Asia. 10,000+ image-question pairs, 11 languages.
- [GG-EZ Regional Adaptation](https://vickyfeliren.com/usecases/gg-ez-regional-adaptation/): (Under Review). SDXL fine-tuning for Southeast Asian cultural adaptation with linear weight-merging. Cultural correctness 1.569 vs 1.491 baseline; 98%+ global benchmark retention.
- [CommonLID Language Identification](https://vickyfeliren.com/usecases/commonlid-language-identification/): ACL 2026. Re-benchmarking LID systems on realistic SEA web data, exposing accuracy gaps on code-switched, romanized, and dialect-heavy text.

### Production ML, Industry
- [Share of Voice Forecasting](https://vickyfeliren.com/usecases/share-of-voice-forecasting/): Artefact (2025). XGBoost + split conformal prediction for Fortune 500 brand SoV forecasting across 6 APAC markets. Calibrated prediction intervals with 90% empirical coverage target.
- [Biometric Auth & Credit Scoring](https://vickyfeliren.com/usecases/biometric-authentication-credit-scoring/): GDP Labs (2021–2023). MobileNet/OpenVINO biometric stack at 1M+ daily inferences, 99.99% uptime. OJK-compliant alternative credit scoring; 35% loan approval efficiency improvement.
- [Fraud Detection Pipeline](https://vickyfeliren.com/usecases/fraud-detection-pipeline/): GDP Labs (2021–2023). Multi-stage pipeline combining rule-based pre-filters, XGBoost, and graph anomaly detection with Kafka streaming and a Redis feature store.
- [Municipal Waste Forecasting](https://vickyfeliren.com/usecases/municipal-waste-forecasting/): Jakarta Smart City (2021). Facebook Prophet ensemble for 10M-resident waste logistics; causal DiD analysis of plastic bag ban. 15% operational efficiency improvement. IEEE ICISS 2021.
- [Demand Forecasting, Consulting](https://vickyfeliren.com/usecases/demand-forecasting-consulting/): Artefact (2025). Audience reach and transaction volume forecasting for Fortune 500 media and financial clients under 2-week consulting cadence.

### Applied & Hackathon Projects
- [HakkTaxi, Ride-Share Demand](https://vickyfeliren.com/usecases/hakktaxi-ride-share/): Microsoft Azure APAC Regional Champion (2020). XGBoost + H3 geospatial grid demand heatmap for Jakarta; 6-second ETA margin of error. Built in 48 hours.
- [TeleHealthMonitor, Edge AI](https://vickyfeliren.com/usecases/telehealthmonitor-edge-ai/): Cambridge CamvsCovid Top 3 globally (2020). On-device respiratory rate estimation for COVID-19 home monitoring on 2G-compatible, GDPR-compliant Android.
- [Community IVR, Voice AI](https://vickyfeliren.com/usecases/community-ivr-voice-ai/): UC Berkeley Cal Hacks Best Community Track (2020). Feature phone AI assistant (ASR + knowledge graph + TTS) for offline Indonesian communities, no smartphone required.
- [Plastic Bag Ban Causal Analysis](https://vickyfeliren.com/usecases/plastic-bag-ban-causal-analysis/): IEEE ICISS 2021 (First Author). Difference-in-differences analysis of Jakarta's 2020 plastic bag ban using 100,000+ NLP-classified citizen complaints.
- [Qwen VL Fine-tuning for AI City Challenge 2026](https://vickyfeliren.com/usecases/aicity-qwen-vl/): LoRA fine-tuning Qwen2.5-VL-3B under 14GB GPU constraints for AI City Challenge 2026 Track 2 using LLaMA-Factory + DeepSpeed ZeRO-2.

### Ongoing Projects
- [Conformal Prediction for VLN](https://vickyfeliren.com/usecases/vln-conformal-prediction/): Thesis research. Parameter-free confidence-rescaled conformal score restores calibrated coverage across VLN-DUET, VLN-HAMT, and Recurrent VLN-BERT on R2R and REVERIE; closed-loop help-seeking lifts success from 71% to 91%.
- [llm-d/inference-scheduler](https://vickyfeliren.com/usecases/llm-d-inference-scheduler/): Open-source contribution. Kubernetes-native LLM inference scheduling with Gateway API, Envoy ext-proc, and KV cache-aware routing for vLLM backends.

## Key Research

- **SEA-VL** (ACL 2025, Core Contributor, Data Pipeline Lead): Multicultural vision-language dataset for Southeast Asia, 10,000+ image-question pairs across 11 SEA languages, 100+ annotators. https://aclanthology.org/2025.acl-long.916/
- **ProCANet** (IEEE GRSL 2025, First Author): Progressive cross-attention network for flood segmentation using multispectral satellite imagery, IoU 0.815, F1 0.898 on Sen1Floods11. https://ieeexplore.ieee.org/document/10750225
- **Mining Segmentation** (Remote Sensing of Environment, IF 11.4, 2025, Co-Author): Multi-modal deep learning for global mining footprint delineation, first use of a geospatial foundation model on this task. https://doi.org/10.1016/j.rse.2024.114584
- **CommonLID** (ACL 2026, Contributor): Language identification benchmark for noisy web data across 109 languages. https://arxiv.org/abs/2601.18026
- **GG-EZ** (Under review, Diffusion Arm): Regional adaptation of VLMs for Southeast Asia; 98%+ global benchmark retention. https://arxiv.org/abs/2604.11490
- **Retention Ponds Study** (AQUA Q2, 2025, Co-Author): Mixed-methods urban resilience evaluation through flood policy in South Bandung. https://doi.org/10.2166/aqua.2025.292
- **Plastic Bag Ban** (IEEE ICISS 2021, First Author): Causal policy analysis of Jakarta's plastic bag ban using citizen complaint data. https://ieeexplore.ieee.org/document/9533236/

## Research Notes

Short distillations of papers, with own take (what the work shows and what it misses). Published at the top of the Writings page.

- **Conformal guarantees are only as honest as the exchangeability they assume** (Jun 2026), on "Language Models with Conformal Factuality Guarantees" (Mohri & Hashimoto, ICML 2024): The cleanest demonstration that you can wrap an LM output in a distribution-free correctness guarantee, but the guarantee rides on exchangeability, which fails in the agentic, multimodal, free-form, low-resource settings that matter. https://arxiv.org/abs/2402.10978
- **We can steer a model toward honesty. We have only checked that in English** (Jun 2026), on "Representation Engineering: A Top-Down Approach to AI Transparency" (Zou et al., 2023): Honesty and harmlessness live in readable, steerable directions and this is becoming default safety tooling, but it is validated almost entirely in English and we do not know whether the probes and steering vectors transfer to other languages or fail silently. https://arxiv.org/abs/2310.01405

## Press Features

- [SEA's AI law trap: who really pays?](https://www.techinasia.com/seas-ai-law-trap-pays) (Tech in Asia, Jun 2026): Featured contributor on who bears the real cost as Southeast Asian governments race to regulate AI.
- [AI rules in SEA: the risks, the fines, what you need to know](https://www.techinasia.com/ai-rules-sea-risks-fines) (Tech in Asia, Jun 2026, paywalled): Featured coverage on AI regulation across Southeast Asia, emerging rules, enforcement risks, and compliance requirements.
- [AI rules in South-East Asia: Risks, fines, and everything else you need to know](https://www.businesstimes.com.sg/international/asean/ai-rules-south-east-asia-risks-fines-and-everything-else-you-need-know) (Business Times, Jun 2026): Same story, open access, AI regulation landscape across Southeast Asia, enforcement risks, and compliance requirements.

## Talks & Public Speaking

- [Do Multilingual Vision-Language Models Abstain under Cross-Modal Conflict in Low-Resource Languages?](https://luma.com/oh51jw9e) (AI Safety India Community Events, Hackathon Winners Present, Jul 2026): Presented the Asia-Pacific winning submission from Apart Research's Global South AI Safety Hackathon.
- PyPalu, Sulawesi Tengah, Indonesia (May 2026): Python for localized context, Python community meetup.
- [MUSE, How Data Science Differs in Each Sector](https://www.monash.edu/indonesia/students/muse/how-data-science-differs-in-each-sector) (Monash University Indonesia, 2025): Panel speaker on how data science practice differs across industry sectors.
- Bank of Indonesia (Oct 2024): Invited talk on data synthesis, data privacy, and responsible data management.
- Bina Nusantara University (Nov 2024): Guest lecture on computer vision, from CNNs to Transformers through semantic segmentation.
- Available Q3 2026 for conference talks, podcasts, and panels on trustworthy AI, conformal prediction, and AI for Southeast Asia.

## Awards & Recognition

- Microsoft Azure Virtual Hackathon APAC, Regional Champion (2020): APAC-wide, Microsoft-featured.
- CamvsCovid, University of Cambridge, Top 3 Globally (2020): Built TeleHealthMonitor for COVID-19 remote patient monitoring.
- Cal Hacks 8.0, UC Berkeley, Best Community Track (2020): Built community IVR voice AI for offline Indonesian communities.
- Reboot the Earth, UN Technology Innovation Labs, Top 10 (2019): Global climate and sustainability solutions challenge.
- IISF 2024, Most Visionary Research: AI and remote sensing for biodiversity impact from energy transition mining.
- Asia Pacific Regional Winner, Global South AI Safety Hackathon (Apart Research, 2026): Multilingual VLM abstention study; presented at AI Safety India's Hackathon Winners event, Jul 2026.
- Monash Indonesia Inaugural Welcome Scholarship (2024): Awarded for research potential in regional AI.
- Deep Learning Nanodegree, Facebook & Udacity, Scholarship (2018).
- Jeffrey Cheah Entrance Scholarship, Sunway Education Group (2015).
- Monash Hackathon, Top 3 (2019).
- Monash University Coding Competition, Honorable Mention (2018).

## Education

- Monash University, Master of Data Science (Expected Sept 2026, GPA 4.0/4.0). Thesis: conformal prediction for vision-language navigation.
- BlueDot Impact, Technical AI Safety (Jun 2026): Cohort intensive covering alignment, interpretability, red-teaming, AI control.
- Monash University, Bachelor of Computer Science (Dec 2019).
- Nanyang Technological University, Computer Science Exchange Programme, Singapore (Jun–Dec 2018).
- Udacity, Deep Learning Nanodegree (2019, Facebook AI scholarship recipient).
- Monash CURIE Compass, Mentee, Centre for Undergraduate Research Initiatives and Excellence (Mar–Dec 2019).

## Writings

- [Knowing when you don't know is the core safety property](https://vickyfeliren.com/essays/knowing-when-you-dont-know/) (Jun 2026, Essay): Why calibrated abstention, not raw capability, is the core safety property of a deployed model; ties first-principles philosophy to conformal prediction, eliciting latent knowledge, and multilingual safety.
- [The cost of becoming](https://medium.com/illumination/the-cost-of-becoming-78f7cd818fda) (May 2026, Personal): A poem for people tired of waiting for themselves.
- [Four hours, one vendor, one airline](https://medium.com/illumination/four-hours-one-vendor-one-airline-1c0044b938c4) (May 2026, Engineering): What a four-hour American Airlines outage reveals about vendor dependency in modern infrastructure.
- [If everyone forgets, why be good](https://medium.com/write-a-catalyst/if-everyone-forgets-why-be-good-0858bdac1bc2) (May 2026, Personal): On moral action when no one remembers, Mother Teresa, Aristotle, and what we owe the universe.
- [Quality and Reliability for AI Engineers](https://medium.com/data-science-collective/quality-and-reliability-for-ai-engineers-b2f92f6406f8) (May 2026, Engineering): Practical guide to evaluation frameworks, failure modes, and engineering practices for trustworthy production AI.
- [We all think our map is the territory](https://medium.com/illumination/we-all-think-our-map-is-the-territory-32b2e50848db) (May 2026, Research): On ethnocentrism as invisible cultural default and building global teams that hold multiple maps at once.
- [Your model just killed someone's grandmother](https://medium.com/technology-hits/your-model-just-killed-someones-grandmother-why-production-ml-system-needs-conformal-prediction-b05245e34ca4) (Jan 2026, Engineering): Why production ML needs conformal prediction, honest uncertainty quantification with mathematical coverage guarantees.
- [The art of letting go while still caring](https://medium.com/illumination/the-art-of-letting-go-while-still-caring-d8e95ca5b063) (Jan 2026, Personal): Personal reflection on control and release.
- [Meditation. 7 days. No phone. Noble silence.](https://medium.com/illumination/meditation-7-days-no-phone-noble-silence-ba3c8da021b3) (Dec 2025, Personal): A data scientist's experience at Taman Brahma Bali Usada.
- [My initial exploration on Trustworthy AI](https://medium.com/technology-hits/my-initial-exploration-on-trustworthy-ai-76f6139f8956) (Nov 2025, Research): Bridging information theory and responsible ML.
- [Phoenix Protocol: Rising from the ashes of adversity](https://medium.com/illumination/phoenix-protocol-rising-from-the-ashes-of-adversity-a5968e079638) (May 2025, Personal): On resilience and reinvention.
- [Privacy Matters: Navigating the Digital Age with Confidence](https://medium.com/technology-hits/privacy-matters-navigating-the-digital-age-with-confidence-6a47dede68ed) (Oct 2024, Engineering): Data synthesis and responsible data stewardship.
- [Mental Health Matters: Please Take Care](https://feliren.medium.com/mental-health-matters-please-take-care-2469d7c62baa) (Jul 2024, Personal): Personal reflection on mental health.
- [A Stranger's Guide to the IU Concert](https://medium.com/illumination/a-strangers-guide-to-the-iu-concert-reflections-on-fandom-culture-and-music-702ec9f03966) (May 2024, Personal): Reflections on fandom and culture.
- [Gone from social media for 4 years. Now I am back](https://feliren.medium.com/gone-from-social-media-for-4-years-now-i-am-back-b835759639d3) (Mar 2024, Personal): Reflections on digital presence.

## Profiles

- Google Scholar: https://scholar.google.com/citations?user=R2LVQ7AAAAAJ
- ORCID: https://orcid.org/0000-0003-3306-8426
- ACL Anthology: https://aclanthology.org/people/vicky-feliren/
- Semantic Scholar: https://www.semanticscholar.org/author/Vicky-Feliren/2330264544
- DBLP: https://dblp.org/pid/392/6206.html
- LinkedIn: https://www.linkedin.com/in/feliren/
- GitHub: https://github.com/feliren88
- HuggingFace: https://huggingface.co/feliren
- Medium: https://medium.com/@feliren
- Devpost: https://devpost.com/Feliren

## Skills

- category: 'RESEARCH: TRUSTWORTHY AI'
  description: 'Knowing when a model should defer, and proving it. Distribution-free guarantees and reliability for deployed AI, beyond calibration scores.'
  skills:
    - 'Conformal Prediction'
    - 'Uncertainty Quantification'
    - 'Vision-Language Models'
    - 'Interpretability'
    - 'Knowledge Distillation'
    - 'Causal Inference'
    - 'A/B Testing'
    - 'Hypothesis Testing'

- category: 'RESEARCH: MULTILINGUAL & MULTIMODAL AI'
  description: 'Making reliability and safety hold for the languages and contexts AI was never built for. Probing and steering internal states, multilingual benchmarks, adaptation methods, and earth-observation models that pressure-test the work against messy data.'
  skills:
    - 'Activation Steering'
    - 'Linear Probing'
    - 'Cultural Benchmarking'
    - 'Vision-Language Adaptation'
    - 'Semantic Segmentation'
    - 'Multispectral Imaging'
    - 'Google Earth Engine'

- category: 'ENGINEERING: DEEP LEARNING & GENERATIVE AI'
  description: 'Model development from architecture design to deployment, multimodal, generative, and earth-observation systems.'
  skills:
    - 'PyTorch'
    - 'TensorFlow'
    - 'scikit-learn'
    - 'LangChain / LangGraph'
    - 'RAG & Agentic Workflows'
    - 'OpenCV'
    - 'Diffusion Models'
    - 'Intel OpenVINO'

- category: 'ENGINEERING: LLM EVALUATION & OBSERVABILITY'
  description: 'Evaluation design and production monitoring, because a model is only as trustworthy as your ability to measure it.'
  skills:
    - 'LLM-as-a-Judge'
    - 'Human-in-the-Loop Evaluation'
    - 'LangSmith'
    - 'MLflow'
    - 'Weights & Biases'
    - 'Calibration Drift Monitoring'
    - 'Latency Benchmarking'

- category: 'ENGINEERING: PRODUCTION ML & DATA'
  description: 'Cloud-scale ML pipelines with distributed inference, designed and operated across GCP and AWS.'
  skills:
    - 'GCP (Vertex AI, BigQuery)'
    - 'AWS (EC2, S3, SageMaker)'
    - 'Kubernetes'
    - 'Docker'
    - 'Apache Spark'
    - 'dbt'
    - 'Vector Database'

- category: 'VISUALIZATION & ANALYTICS'
  description: 'Translating model outputs and data findings into decisions, for technical and non-technical audiences alike.'
  skills:
    - 'Tableau'
    - 'Streamlit'
    - 'Plotly'
    - 'Matplotlib'
    - 'd3'

- category: 'LANGUAGES & FOUNDATIONS'
  description: 'Core tooling for research and systems engineering.'
  skills:
    - 'Python'
    - 'Scala'
    - 'SQL'
    - 'R'
    - 'Git'
    - 'Linux/Unix'