Anand Umashankar (Anand CU)

ThreeV.ai's platform turns drone and helicopter imagery of power lines into inspection findings for US utilities. I lead ML across modelling, MLOps, data operations and labelling, reporting to the CTO and working remotely from Bengaluru with a team across the US and Europe, and take on work beyond the ML remit: product scoping, UX, hiring and client advisory.

• Architected the agentic inspection system now in production with US utility customers.
• Scoped, designed and shipped the evaluation and review app largely myself with coding agents.
• Set up the model-serving platform and the MLOps stack: MLflow tracking and registry, DVC-versioned datasets, CVAT and Voxel51, accuracy-gated releases, automated model cards.
• Led the CV models behind the platform: MaskDINO and Mask2Former instance segmentation on high-resolution drone imagery, a Qwen3-VL LoRA defect classifier and monocular-depth measurement. Built the depth-based 3D multi-view asset counting and wire triangulation in production.
• Designed the asset ontology and attribute labelling taxonomy end to end, from spec to trained model: component classes with material, configuration and position attributes mapped from utility inspection standards and job aids; an attribute-based classification framework that turns inspector requirements into attribute heads on the segmentation models; labelling guides, vendor labelling and QA; evaluation per attribute. A job-aid gap analysis found inspection questions the platform was missing and added them.
• Built, solo, "Ask Utility Inspector", an internal RAG assistant used by the team over industry textbooks, maintenance manuals, job aids and labelling guides: heading-aware text chunks in pgvector plus ColQwen page-image embeddings (hybrid retrieval), answers from the Claude API.
• Built the team's single ML knowledge base: inspection-question priorities, the use-case roadmap, labelling guides and model cards in one repo, synced daily by GitHub Actions; labelling guides and model cards are written as source and compiled and published from CI.
• Supported client benchmark and fine-tuning contests (class-wise thresholds, NMS, metric mapping, SAHI tuning), utility pilots and demos, and a utility RFP narrative and orals. ML resource owner in the company's AWS and GCP cost-reduction review.
• Owned an AI/ML advisory engagement for a large US utility end to end and delivered over half of it: AI/ML data roadmap, image-only (LiDAR-free) measurement feasibility study, ML solution blueprint, MLOps and IT readiness plan, a cross-utility model-sharing platform blueprint and a two-year labelling roadmap.
• Security and engineering hygiene for SOC 2: Trivy and grype scanning in CI, with a scheduled job that uses the Claude API to draft remediation plans.

• Hiring, onboarding with reading lists and 90-day plans, role definitions, performance reviews, one-to-ones.
• Co-wrote the ML operating model: the inspection question as the unit of work, the agentic path first with narrow models only when evaluation justifies them, ground truth before any model, evaluation as the promotion gate.
• Started and ran the team's cadence: daily ML stand-up with bi-weekly sprint consolidation, a weekly prioritisation review, a monthly model review for leadership, sprint review and planning across the ML, MLOps and labelling lines, and use-case-driven quarterly planning. Automated the monthly review and sprint decks with Claude and packaged them as reusable agent skills. Design documents and design reviews.
• Ran a recorded knowledge-transfer programme when a senior ML engineer left, so wire, insulator-defect and labelling context stayed with the team.
• I delegate the core ML work by line (depth experiments and evaluation to data science, labelling specs and ground-truth review to data operations, deployment to MLOps) and use coding agents for the glue: internal apps, tools and prototypes that would otherwise never get headcount.
• AI-native development: Claude Code with parallel agents in git worktrees, plus an "architect" agent that writes daily briefs and weekly updates.

Senior Machine Learning Engineer, Sep 2021 – Nov 2023 → Machine Learning Tech Lead, Dec 2023 – Dec 2024

Sharper Shape does AI analytics for aerial LiDAR and imagery of power lines. As Tech Lead: led 10 across ML and annotation (6 direct reports); model results delivered to major California utilities.

• Drove the move of the LiDAR segmentation model from a graph neural network to a sparse 3D CNN (MinkowskiEngine), backed by new training data from the annotation team, making inference cheap enough to run on every LiDAR file the company processed.
• Went beyond the ML remit: worked with the delivery team to find where the model failed (errors they were cleaning by hand before delivery) and connected the processing tooling to call the model.
• Led the ML behind LiDAR Insights, the company's AI point-cloud classification product and highest-demand ML line: 29 utility classes (8 wire classes) at 95–98% accuracy and F1 up to 0.98, cutting delivery from months to days or a few weeks.
• Scaled the production CV and LiDAR pipelines behind the company's 6.4 PB of aerial data and 42,000 miles (67,000 km) of power lines flown, with 8x faster processing.
• Ran ML from business requirements through research and proofs of concept to production, and helped plan the ML team's OKRs.
• Built multi-GPU training workflows with PyTorch Lightning.

• Achieved state-of-the-art speech recognition with an end-to-end attention-based encoder-decoder.
• Built distributed training with WebDataset over 26,000 hours of audio, among the largest ASR training sets at the time.
• Built the data pipeline for a new ~26,000-hour corpus of business conference calls: forced-aligned it into ~9M utterances from 52,484 speakers, then repacked ~5 TB of small files into 3,000 WebDataset shards (2.1 TB), making data loading 7–8x faster.
• Word error rate 10.87%, against 19.7% (Azure) and 21.9% (Google) cloud speech-to-text on samples of the same data.
• Showed that more data trains faster, not slower: the 20,000-hour model reached the benchmark error in 48 hours vs 76 hours for 8,000 hours. Synchronous 4-GPU data-parallel training (V100s) halved training time and improved accuracy; asynchronous Hogwild did not.
• Thesis: "Large Scale Speech Recognition with Deep Learning" (Aalto University and TU Berlin, Sep 2021; supervisors Prof. Mikko Kurimo and Prof. Sebastian Möller). PyTorch, SpeechBrain.
• Project during the MSc: federated learning for remote-sensing data (TU Berlin, 2020).

• Built a real-time motorcycle driver-assistance system for India: multi-camera CNN detection, object tracking and trajectory estimation.
• Shipped a traffic-management vision system deployed in Argentina on sub-$100 edge hardware, with model quantization to run on it.