Siddharth
Raj.
ML/AI Researcher · MS CS @ University of Chicago · Ex-Walmart Global Tech
I build ML systems that are both theoretically grounded and production-ready. My work spans RLHF alignment pipelines, geometric computer vision with transformers, and high-throughput NLP systems.
I'm a Master's student in Computer Science (AI Specialization) at the University of Chicago, with 3+ years of industry experience as a Software Engineer III at Walmart Global Tech building production LLM systems and distributed backends.
My research interests live at the intersection of geometric computer vision (injecting epipolar priors into transformer matchers), LLM alignment (RLHF pipelines, reward modeling, PPO), and efficient ML systems (LoRA, ONNX deployment, MLOps at scale).
I hold a B.Tech from IIT BHU and am a Candidate Master on Codeforces — strong algorithmic foundations shape how I think about every problem I tackle.
I care about work that bridges the gap between research insight and real-world impact. Whether that's getting an RLHF pipeline to run on a single consumer GPU, or diagnosing a confidence distribution shift in a fine-tuned vision transformer, I like to understand systems deeply.
End-to-end 3-stage alignment pipeline: SFT → Reward Model (Bradley-Terry pairwise ranking, 82% human agreement) → PPO optimization with KL-divergence penalty to prevent reward hacking. LoRA adapters cut trainable params from 6.7B to 33M (0.5%), enabling single-GPU training. W&B tracking throughout.
Injected differentiable epipolar constraints into a transformer-based feature matcher via a soft fundamental-matrix mask on the confidence matrix. Fine-tuned with focal loss + LoRA adapters on ScanNet; diagnosed confidence distribution shift post-training — threshold recalibration to 0.10 recovered 436 matches/pair while preserving geometric precision.
Modeled Uber/Lyft as a Mean-Field Game over 259 NYC TLC zones using real HVFHV trip data. Backward induction + fictitious play finds the Nash equilibrium in 166 iterations — drivers at JFK queue for high-fare Manhattan runs. Interactive map shows driver redistribution and hourly demand across the city.
Custom PyTorch CNN on CIFAR-10 (BatchNorm + residual connections, Adam + StepLR). Exported to ONNX Runtime via FastAPI for production inference; deployed on AWS ECS via ECR. Grad-CAM explainability surfaced systematic animal misclassifications — targeted augmentation boosted accuracy 2.5%. Model and weights published to HuggingFace.
4-stage pipeline over 10,000+ reviews at ~9,100 reviews/sec. Combines VADER lexical sentiment (69.3% star-rating agreement) with BERTopic — sentence transformers + UMAP dimensionality reduction + HDBSCAN clustering — to auto-discover 44 latent topics and extract 10 composite priority pain points. Served via a Flask analytics dashboard.
Multi-threaded ray tracer in Go implementing global illumination, reflections, and shadows. Work-stealing scheduler with lock-free deques distributes tile partitions across 15 cores for load balancing under variable render costs. Profiled on UChicago Peanut cluster with Go's pprof — 12× speedup over sequential baseline.
Built LSCM UV parameterization from scratch using sparse Cauchy-Riemann equations over bmesh topology — mathematically optimal angle-preserving mapping for curved surfaces. Modeled multi-body collision dynamics via quadratic B-Splines with C0 knot multiplicity at impact points. All geometry, shader nodes, and particle systems generated programmatically via the bpy API.
- Shipped a speech-to-text research platform on AWS Amplify (Next.js, Python) enabling researchers to train and deploy custom language models; NIST 800-171 compliance; GraphQL APIs over S3 + DynamoDB; Okta / Grouper / Cognito SSO
- Built multi-level auth and authorization workflows via AWS Cognito + Okta securing 200+ university staff across role-based permission tiers (Inventory Management Platform, DynamoDB)
- Built Python middleware syncing JIRA with SolarWinds via webhooks and CrowdStrike via APIs, automating IT operations event routing
- Developed an internal AI chatbot using GPT-4o, LangGraph, and Milvus (RAG) — reduced seller support tickets by 30%; Python backends + React UI, end-to-end
- Built an Elasticsearch Translation Memory System using KNN + fuzzy search — cut translation costs 30% and enabled 80% localization coverage across seller apps
- Designed a Redis caching layer over MongoDB — dropped P95 API latency from 800ms to under 200ms across 10+ microservices
- Parallelized seller onboarding Java workflows — 36× faster (3 days → 2 hours)
- Designed 10+ Spring Boot microservices with Kafka pipelines (DLQs, zero message loss); mentored 3 junior engineers on production debugging and RAG design
- Presented "WICR: Smarter and Contextual Language Translation Reviews" at Walmart's internal research summit — top 25 of 800 submissions
I'm open to research collaborations, full-time ML/AI roles, and interesting conversations about alignment, geometric vision, and building systems that actually work.