Ansh Mittal

Machine Learning Engineer II

Machine Learning Engineer II @ Guidewire | 5+ yrs ML (3D CV + LLM Optimization). Spanning foundational research to enterprise production. Proven track record of shipping optimized spatial pipelines and architecting large-scale agentic AI systems. Skilled in translating cutting-edge prototypes (NeRFs, Gaussian Splats) and fine-tuned Language models into scalable, production-ready workflows for complex computing challenges.

Contact Me

Career Systems + Hyperbolic ML

Interactive 3D world model

Poincare Job Intelligence

A fact-locked career-market graph that maps verified resume evidence to ML role tracks, S&P 500 and Nasdaq-100 companies, ATS platforms, and discovered openings using deterministic three-dimensional Poincare embeddings.

Human-approved submissionsS&P 500 + Nasdaq-100Workday, Greenhouse, AshbyHQ, Lever, Eightfold, iCIMS

Explore career graph

January 17, 2025

CUDAx365 Animation Set

Black Hole 3D Flow

CUDAx365 black-hole simulations and Neural Radiance Field experiments for dynamic flow analysis, MHD graph kernels, and approximation studies.

3D hot/cold stellar flowMHD CUDA graph kernel combinedMHD CUDA approximations

Quantum Error Correction

View Documentation →

Real-Time QEC Decoder

Surface-code decoder prototype with C++17 runtime, Python transformer path, benchmark drift checks, and CUDA-Q/cuBLAS/cuRAND hooks. C++17 runtime path: p50 0.000433 ms over 1024 shots on the CPU-only runner. Python transformer prototype path: p50 0.003060 ms on the same benchmark setup.

C++17 runtimePython transformer pathCUDA-Q/cuBLAS/cuRAND

WorldSim v0.2

Offline simulator

WorldSim Multi-Field Radio Digital Twin

Offline simulator combining synthetic private-IP host workloads, QFT-inspired phase-space fields, room-aware WiFi/radio propagation, RSSM world-model trajectories, vector-quantized event-window detection, and Wireshark-readable PCAP exports per job.

Synthetic PCAP exportsRoom-aware radio fieldsRSSM + VQ event detection

3D Medical Imaging

Volumetric U-Net

3D CT Scan Segmentation

A U-Net-style segmentation pipeline adapted from 2D slices to volumetric CT data by using 3D feature blocks, skip-connected encoder-decoder paths, and spatial context across neighboring slices.

3D CT volumesEncoder-decoder segmentationSlice-to-slice context

Fundus CAD Research

EfficientNet B1

EfficientNet Fundus Classification

Fundus disease classification across 40 abnormality labels using fine-tuned pretrained CNNs. EfficientNet B1 was the strongest single-model choice, with high precision/recall on the reported split and a practical path toward computer-aided retinal screening.

40 fundus labelsFine-tuned CNNsCAD for retinal screening

World Models

VAE + MDN-RNN + CMA-ES

World Model Implementation

Implementation of the World Models agent pattern: a CNN-VAE compresses image frames, an MDN-RNN predicts future latent states, and CMA-ES optimizes a compact controller over the learned latent memory.

Latent visual modelMixture-density memoryEvolutionary controller

AR + AI Insurance

AR INSURIM

AR Insurance Prototype

An AR insurance prototype for inspecting policy objects and claim context in-place. The same flow can evolve with AI-assisted damage detection, guided evidence capture, and real-time claim triage.

AR inspectionAI claim triageGuided evidence capture