Rijul Dahiya
Forward Deployed AI Engineer, PollyLabs
M.S. Computer Science, NYU Courant & CDS
Jersey City, NJ · [email protected]
I build ML systems that generate, understand and control visual and sequential data. Currently a Forward Deployed AI Engineer at PollyLabs. Previously a graduate research assistant at NYU CDS working with IBM Research on geospatial foundation models, ML Engineer Intern at Agree.com, SDE at Groupon, and research at the University of Washington, Arizona State and ISRO.
I like turning ambiguous problems into production systems: the kind that survive contact with real data, real latency budgets and real users.
$516,124.50
recovered in unpaid guaranteed compensation for laborers in India through a grievance reporting platform I built during the COVID recovery period.
Best Presentation Award
at Hiroshima University for an SDG-focused international research project.
1 of 6 worldwide
selected for an entrepreneurship program run by the Japan Science and Technology Agency and the Japanese government.
Research topics I am currently interested in:
- World models & diffusion: Reward design and post-training for diffusion-transformer world models, fine-tuning them toward controllable, physically coherent rollouts without paired ground truth.
- Mechanistic interpretability: Reading and writing model internals: latent activation extraction and inference-time patching to steer large vision-language models without touching their weights.
- Image editing & generation: Instruction-guided editing that preserves what it should preserve, using seed selection and dual-objective scoring for faithful, localised edits.
- LLMs, RAG & offline RL: Retrieval-augmented systems in production, and diffusion planners for decision-making where the reward is sparse and the horizon is long.
Education
New York University, M.S. Computer Science (AI concentration) · Courant Institute of Mathematical Sciences · New York
Coursework: Machine Learning, Deep Learning, NLP, Computer Vision, Reinforcement Learning, LLVM, Big Data & ML Systems.
BITS Pilani, Hyderabad, B.E. Computer Science (Minor in Data Science) · Birla Institute of Technology and Science · Hyderabad
Merit Scholarship.
Experience
Software and ML systems I have shipped in industry.
PollyLabs, Forward Deployed AI Engineer · New York, US
Agree.com, ML Engineer Intern · New York, US
Trained and shipped a production PDF object-detection pipeline with LLM labeling, tracking experiments in MLflow and Weights & Biases to reach 90.6% accuracy. Built a RAG system over a Spark-processed contract corpus for payment-intent extraction, cutting review time ~70% and latency 65%. Deployed containerised inference (Docker, LangChain, ONNX) with model monitoring for +30% throughput.
Groupon, Software Development Engineer · Remote, India
Owned scalable customer-facing email infrastructure (React, FTL) delivering 1M+ emails across APAC and EMEA and 12+ markets. Built and deployed a bulk-campaign service (Node.js, Express, PostgreSQL) cutting creation time ~70%, and led an internal transactional email system supporting an 18% month-over-month rise in campaign output.
Bank of New York Pershing, SDE Intern · Chennai, India
Implemented RBAC-secured CRUD workflows and query filters for the NetX360 custody module serving 100K+ RIAs across ~8M accounts. Integrated a LangChain/LangGraph RAG chatbot improving data retrieval ~30% for 5,000 Pershing users, and built full-stack features in Java Spring Boot, Angular and SQL.
Providence Healthcare, SDE Intern · Remote
Deployed a React/.NET authentication framework with SQL Server, securing healthcare data across 20+ applications for 5,000+ medical professionals.
ISRO, Research Intern · India
Engineered LiDAR ML pipelines over 500+ km² datasets generating high-resolution Digital Elevation Models, improving mapping accuracy 12% and runtime 35%.
Research
Labs and collaborations, most recent first.
NYU Center for Data Science, Graduate Research Assistant · New York
Worked with IBM Research on SATML model benchmarks; contributed to TorchGeo and TerraTorch. TA for ML for Climate Change and Programming Tools for Data Scientists.
IBM Research, Research Collaboration with NYU · New York
Developed multi-temporal latent-space fusion in TerraTorch/TorchGeo (PyTorch), lifting substation detection 15%.
Arizona State University, Summer Research Initiative · Tempe, AZ
Optimised Logan River streamflow estimation by 8.33% using a channel- and spatial-attention ConvLSTM (NSE 0.846, PBIAS 0.038% on the WY2009–WY2018 test period).
University of Washington, Thesis · Prof. Richard Anderson, ICTD Lab · Seattle, WA
Built a reproducible data platform for the eKichabi app, categorising 12 distinct user-behaviour patterns from financial transaction data via unsupervised learning.
Hiroshima University, Sakura Science Fellowship · Hiroshima, Japan
Analysed sustainability challenges via ML; won the Best Presentation Award.
Projects
Papers and systems. Each links to the full write-up, figures and PDF.
GameWorldScore: Ground-Truth-Free World Model Fine-Tuning
A unified five-signal reward framework and GRPO fine-tuning pipeline for diffusion-transformer world models. The unified model matches every single-reward specialist on its own metric while holding the rest (action fidelity 0.469 vs 0.466, anti-drift 0.842 vs 0.859).
GRPODiffusionWorld ModelsDO-ELECT: Instruction-Guided Image Editing
Dual-objective seed selection that balances preservation against change: SSIM 0.664 → 0.839 on MagicBrush with strict compositing, at a CLIP cost under 0.008.
DiffusionImage EditingPatch and Control: Steering LVLM Behaviour
Steered LLaVA-1.5 via latent activation extraction and inference-time patching; early decoder-layer patching gave the highest behaviour-transfer fidelity (BLEU, BERTScore, COMET), enabling control without fine-tuning.
InterpretabilityLVLMSLURMDiffuser++: Velocity Prediction & MPPI for RL
A velocity-prediction diffusion planner with MPPI and replanning: 21.1% over Diffuser on U-Maze and 14.5% on Medium, at roughly half the planning horizon.
Offline RLDiffusionMPPIAudio Classification: BEATs on AudioSet
Self-supervised BEATs features against eleven classifiers on AudioSet: handcrafted features fail to converge, and last-layer embeddings reach F1 0.514 with a GRU.
AudioSSLLung Sounds: ViT + CNN
Vision transformers over STFT, CWT and Stockwell spectrograms; 90% test accuracy on normal versus pneumonia.
HealthcareViTStreamflow Prediction: Channel-Attention ConvLSTM
NSE 0.846 on the Logan River watershed; 8.33% improvement over prior models.
ClimateConvLSTMeKichabi Dashboard: Dual-Platform Agricultural Tech
A USSD + Android directory for 9,833 agricultural businesses in rural Tanzania.
ICTDMobileTrading Bots Platform: Cloud Stock-Market Simulation
A multi-language bot competition platform with containerised execution and dynamic scoring.
CloudDockerBug Prioritization: Hierarchical Attention Networks
Hierarchical attention over Linux and Chromium bug reports; concatenating an emotion score beat every embedding swap, +0.038 F1.
NLPHAN
Building
Systems that shipped rather than papers that were written. The repository is the artifact, so each links straight to source.
A logistics exception intelligence system: a map-driven operations console that watches a live fleet, intersects National Weather Service alerts with route geometry, and runs a pipeline of six IBM watsonx Orchestrate agents when a shipment is about to break. The agents trigger the exception, decide fulfilment swap versus reroute versus hold against nearby inventory, rank route alternatives while rejecting premiums above the SLA penalty, draft the driver and customer notice, and synthesise an executive verdict. Every agent returns a single strict JSON object so the UI can render it structurally.
Next.js 16React 19TypeScriptwatsonx OrchestrateRedisGoogle MapsAgent Surf: One-click internet access for AI agents
Connecting an agent to Twitter, Reddit, YouTube, Hacker News, Product Hunt or StockTwits means finding, installing and debugging a different access path for each, every one with its own barrier — metered APIs, blocked server IPs, developer tokens. Agent Surf picks the most reliable path per platform, installs it and health-checks it, so backend churn stays invisible to the agent calling it. MIT licensed.
Python 3.10+MCPMIT licenceAnchor: BPD recovery companion, agentic mesh
A React Native mental-health app supporting three roles — patient, ally and therapist — with DBT-informed tools, mood tracking and safety contracts, backed by five specialised AI agents. Built with an explicit safety posture: a security audit up front, and a standing disclaimer that it is a prototype, not HIPAA-compliant, not FDA-cleared, and no substitute for professional care.
React NativeExpo 54Express 5MongoDBDigitalOcean Gradient AIBedtime Story Pipeline: Generate → judge → revise
A multi-agent content pipeline built around three principles, each enforced by a mechanism rather than a prompt wish: safety as a hard gate, a descending emotional arc that reduces tension in the final third, and read-aloud quality. The judge is deliberately two-tier — hard blockers trigger full regeneration, because unsafe content is never patched, while a soft six-dimension rubric triggers targeted revision. Agents are separated so each stage can run at its own temperature: deterministic for classification and judging, creative for the draft.
PythonOpenAI APIMulti-agent orchestrationYuka: Food label scanner and nutrition scorer
Photograph a packaged-food label and get a 0–100 nutrition score plus the ingredients worth knowing about. Three stages: on-device OCR with Tesseract.js, then Gemini to pull the ingredient list out of noisy OCR text, then Gemini again to score it. The scoring prompt bans vague hedging outright — a flagged ingredient has to name a real mechanism or it gets dropped.
React 19Vite 8Tesseract.jsGemini APITypeScriptHuman-AI Choreography: Laban movement analysis for dance
Semi-supervised learning across motion capture and text, mapping dance movement to and from natural-language description through Laban effort qualities. Includes a Dash-based labelling tool for annotating motion-capture sequences and visualisation tooling for rendering sequences as animations.
PyTorchTransformersDashMotion captureW&BDO-ELECT: Dual-objective seed selection, reference implementation
The released implementation behind the PreserveChange paper: dual relevance maps, Dual-BIS scoring and a strict compositing mode, packaged as a CLI over InstructPix2Pix and MagicBrush backbones. Extends and credits the original ELECT codebase.
PythonPyTorchDiffusersCUDAGameWorldScore: GRPO fine-tuning stack for world models
Around 8,000 lines of training infrastructure behind the GameWorldScore paper: five reward modules, a GRPO trainer with group-normalised advantages, actor and rollout workers, Minecraft dataset loaders, SLURM launch scripts and checkpoint evaluation tooling.
PyTorchGRPORaySLURMW&BPeri-Urban Hyderabad Change Detection: Open remote-sensing dataset
A published dataset of 56 GeoTIFF images covering peri-urban Hyderabad: surface reflectance for 2019, 2020 and 2021 plus a seven-band pollution image per area (NO₂, SO₂, HCHO, CO, O₃, AER_AI, CH₄), assembled for change-detection research.
GeoTIFFRemote sensingGoogle Earth EngineMosaic: Three-app Django platform
A movie finder with Markdown editing, calendar integration and Chart.js popularity graphs, a discussion board, and a marketplace for movie memorabilia — three interconnected Django applications sharing Google OAuth, pagination and route protection.
DjangoPythonJavaScriptChart.jsGoogle OAuthSwap: Campus marketplace, Android + backend
A student marketplace shipped as a paired product: a Python backend deployed to Heroku with institutional authentication, and a Kotlin Android client.
PythonKotlinAndroidHerokuLung Sound App: Point-of-care companion to the ViT classifier
A React-based mobile front end for the lung-sound research: record breathing through the phone microphone, run the trained model, and return a classification with guidance toward further testing.
ReactMobileTensorFlow
Open Source
Libraries I have contributed to through research collaborations.
- TerraTorch
Multi-temporal latent-space fusion for geospatial foundation models, developed during the NYU CDS collaboration with IBM Research.
- TorchGeo
Datasets, samplers, transforms and pre-trained models for geospatial data; contributions alongside SATML benchmark work at NYU CDS.
Technical Skills
PyTorch, TensorFlow, Scikit-learn, Deep Learning, NLP, Computer Vision, Gen AI, LLMs, Transformers, RAG
Docker, Kubernetes, CI/CD, MLflow, Weights & Biases, Model Monitoring, ONNX
Spark, Ray, Kafka, Databricks, REST APIs, GCP, Azure
Python, Java, C++, C, SQL, JavaScript, TypeScript, Elixir
React, Angular, Node.js, Express, FastAPI, Spring Boot, LangChain, LangGraph
Writing
Notes on AI, engineering and building systems, published on Medium.
Read the articlesContact
Open to research collaborations and roles in applied ML. Email is the fastest way to reach me.