Vignesh Saptarishi

Machine Learning Systems · Research, Algorithms & Engineering

vignesh.saptarishi@gmail.com | saptaxis.dev | github.com/saptaxis
linkedin.com/in/vignesh-saptarishi | traitful.ai

10+ years engineering production AI systems across computer vision, NLP, search and retrieval, generative models, agentic workflows, and ML platforms. Zero-to-one algorithms engineer on three revenue-generating products, subsequently building and leading the ML and data organizations behind them. Technical lead in enterprise sales to major clients including Macy’s, Levi’s, FedEx, and HDFC Bank. Computational neuroscience research at Brandeis on an NIH-funded PhD fellowship; electrical engineering by degree.

Professional Experience

Feb 2025 - Present AI/ML Engineering Practice · Traitful AI Labs

Independent engineering practice building domain-specific AI systems for clients, alongside original product development and research. Engagements across logistics, legal, and finance, from prototype to production.

Logistics - agentic search and document intelligence

  • engineered production hybrid search and trusted write paths across deeply nested structured records and unstructured documents: domain-specific taxonomies, lexical/vector search in Elasticsearch, LLM-generated queries, and Terraform deployment
  • built a custom agent runtime over raw model APIs, explicitly assembling per-turn context, state, tools, and prompts to bound latency and token costs as conversations grow
  • developed programmable agent memory exposed through model-controlled tools, with runtime-selected agent topologies, tools, and human approval gates
  • built a scenario-based evaluation harness with miniature simulated worlds, grading retrieval, write-path safety, idempotency, and evidence recall, including checks for decisive evidence in memory
  • fine-tuned ModernBERT with a shared encoder and dozens of classification heads to label semantic sections across document types, driving retrieval, extraction, and prompt-example selection

Other engagements

  • financial compliance automation with LLM workflows around deterministic, verifiable calculations
  • desktop automation for embassies through the accessibility tree

Jan 2020 - Oct 2024 Director - Machine Learning · vue.ai / Mad Street Den

Built and led ML and data organizations across three products, each with its own team and stakeholders; wrote production code through the Director role.

Enterprise AI Platform

  • one of the key architects of the company’s unified AI platform: four solution hubs over a shared data layer, an ML studio and DAG-based orchestration
  • designed shared platform abstractions derived from production patterns across algorithms, data models, workflows, and inference
  • enabled non-technical users to train models, compose workflows, and map taxonomies through a no-code interface

Intelligent Document Processing

  • zero-to-one founding algorithms engineer; led ML and product for a domain-agnostic document processing platform, the company’s fastest product to seven-figure ARR, now Automation Hub
  • built the full stack from OCR to structured extraction: table/figure detection, custom embeddings, re-training loops, and hybrid layout understanding combining LayoutLM with geometric heuristics
  • one-shot document onboarding via hybrid text and image registration, enabling extraction from new document types with a single annotated sample and no retraining
  • developed early retrieval-augmented extraction workflows, navigating constrained context windows and client taxonomies spanning thousands of attributes.
  • engineered pre-LLM-era pipelines processing millions of documents with profitable unit economics, optimizing straight-through-processing with customer-validated precision-confidence thresholds
  • deployed across government (Dubai CommerCity), BFSI (HDFC Bank, Sundaram Finance), and legal sectors
  • stack: BERT, LayoutLM, Hugging Face Transformers; NER, classification, embeddings/search, document layout analysis, computer vision, multimodal extraction, RAG, and in-context learning

Virtual Dressing Room

  • pioneered GAN-based virtual try-on, among the earliest production applications of GANs in this space
  • reframed generation as geometric deformation, bridging 512px model resolution to photorealistic 4K output by regressing and upsampling warp grids, then sampling from original images to avoid generative artifacts
  • designed and trained core networks for garment transfer, lighting and shadow synthesis; enabled quadratic growth in outfit combinations with only linear preparation effort
  • built a large-scale synthetic data pipeline using Blender for 3D modeling, CLO3D/Marvelous Designer for garment physics, and automated rendering across hundreds of machines
  • engineered human-in-the-loop production workflows with Photoshop-native mesh outputs, separate lighting/shadow layers, in-app inference plugins, and corrections fed back into training; reduced retouching time per garment from 30–60+ to 2–10 minutes
  • led a six-person algorithms research team and set technical direction for a 50+ person production and data organization reporting into Product
  • stack: conditional GANs, spatial transformers, dense warp fields, segmentation, garment/body pose detection, lighting and shadow synthesis; PyTorch, Blender, CLO3D

Feb 2017 - Jan 2020 Senior ML Engineer & Technical Lead · vue.ai / Mad Street Den

Automated Product Tagging for eCommerce

  • founding algorithms engineer; delivered production computer vision for enterprise retail and fashion clients, including Fortune 500 companies; customers included Macy’s, Levi’s, Diesel, and ThredUp; now the platform’s Data Hub
  • built a unified fashion taxonomy with per-client label mappings, enabling new-client deployment in a single working day by non-technical staff, with ML support only for exceptions; processed hundreds of millions of catalogue images
  • designed taxonomy-aware confidence thresholds and refined the classification hierarchy using confusion matrices
  • architected single-pass feature extraction with lightweight classification heads serving hundreds of models from a shared fleet, with automated retraining and no-code model building
  • built and led a ten-person data annotation team, producing hundreds of thousands of labelled images
  • stack: fine-grained hierarchical classification, object detection, garment/body segmentation, video scene understanding, cross-modal image–text search, generative product copy; Chainer, PyTorch, Inception, ResNet, Faster R-CNN, YOLO

2010 - 2015 Computational Neuroscience Research

Marder Lab, Brandeis University, 2013–2015

  • how a neuron keeps its behavior stable while its own geometry changes underneath it, modeled as coupled nonlinear dynamics with feedback control on growing spatial structures, and studied in simulation

Waran Research Foundation, India, 2010–2013, undergraduate research

Projects

  • built a parameterized Lunar Lander with seven configurable physics parameters; trained PPO controllers with and without explicit physics access, finding that physics-aware agents learned smoother control but often generalized worse under distribution shift (latent-world-physics)
  • world-model-ladder: built a progressive world-model training framework spanning linear models, MLPs, GRUs and RSSMs, comparing single-step prediction, multi-step rollout, scheduled sampling and variational objectives under shared training and evaluation protocols
  • world-model-audits: built a representation and dynamics audit suite for learned world models, testing latent-state decodability, predictive value beyond identity baselines, and action-conditioned physical consistency; applied to published JEPA and RSSM architectures
  • scad and orglens: built persistent work tracking for coding agents; scad preserves searchable session histories, durable memos and cross-session handoffs; orglens models units of work spanning repositories, documents and agent sessions, deriving status and activity from their records
  • The Wooster Papers: a fan site for the Jeeves and Wooster books, with every scene indexed by its characters, places and incidents; the indexing and the search are model-assisted, so a half-remembered episode can still be found

Education

Aug 2013 - Aug 2015 M.S. Neuroscience - Brandeis University, USA

  • graduate researcher in Prof. Eve Marder’s lab on an NIH fellowship
  • graduate coursework in applied dynamical systems, mathematics and molecular biology; GPA 3.84/4.00
  • teaching assistant, undergraduate behavioral neuroscience with Prof. Don Katz

Sep 2009 - May 2013 B.Engg. Electrical Engineering - Anna University, India

  • coursework in control theory, numerical methods and signal processing; GPA 8.0/10.0
  • thesis: Methods for Nonlinear Estimation using Kalman Filters to estimate state vectors and control DC motors

Fellowships & Awards

  • National Institute of Health - Neuroscience PhD Fellowship, Brandeis University, 2013-2015
  • Chairman (Devaki Muthiah) Endowment Award for Overall Best Student 2009 - 2013, Sri Venkateswara College of Engineering, Anna University