Vijay Sojitra

Agentic AI Architect & Consultant

Agentic AI Operations • LLMOps • MLOps • Data Science • AI Engineering/Infrastructure

vjsojitra8@gmail.com

Takes AI from raw data to dependable production systems: governed pipelines, rigorous statistical modeling, evaluated ML and LLM deployments, and the APIs, monitoring, and decision tools that teams rely on every day.

10 years Enterprise AI Systems Design, Machine Learning, Data Engineering & Science and Business Analytics
Dual Master’s Degrees Statistics and Statistical Data Science
Selected Work Contact

INDEPENDENT ENGINEERING

Independently built systems, each with tests, evaluation, and documented trade-offs. They span agentic AI, document intelligence, data products, serverless cloud pipelines, and open-source contribution.

Featured Project

Agentic Document AI

BillFlow: Agentic Utility-Bill Retrieval, Extraction & Payment

An end-to-end system that signs in to utility portals, retrieves new bills, extracts source-grounded fields, and sends payments through a deterministic policy gate. Read-only retrieval and irreversible payment run as separate trust domains.

Measured: 96.3% field-level F1 (exact match) on native-PDF bills, scored by a 105-document evaluation harness with a CI regression gate.

  • LangGraph retrieval agent with Postgres checkpointing, OTP/MFA handling, integrity checks, and resumable runs. It uses LLM self-healing selector repair against adversarial mock portals that inject 500s, 429s, layout changes, and corrupt PDFs.
  • Tiered extraction: text layer, issuer classification, layout OCR, then Azure OpenAI structured outputs. Every field must cite a source span that is verified to exist in the document, then pass arithmetic and meter-read validation.
  • Streamlit review queue with PDF crops and validated corrections, plus a dry-run payment graph with payee allowlists, idempotency, an anomaly ceiling, and a hash-verified audit trail
  • Six synthetic degradations (scan, skew, JPEG, phone photo, stamp) per source bill, LangSmith tracing, and 110 automated tests

Private repository. Architecture walkthrough and end-to-end demo available.

Data Product & Statistical Modeling

ixo8: SaaS Revenue Intelligence with Published Model Error

A full-stack platform that models SaaS revenue from public signals, then scores those estimates against revenue verified from the companies' own payment processors.

  • Company discovery from Meta Ad Library sweeps, qualification filters, and about 10 parallel enrichment connectors. Every datapoint is stored with its source URL and timestamp.
  • Read-only connectors for nine payment processors (Stripe, Paddle, Polar, Lemon Squeezy and others) plus GA4 and Search Console, with encrypted owner credentials
  • Calibration is blocked unless it passes four acceptance gates. Estimate ranges are fit on a held-out split of an 818-company truth set, raising measured coverage from 14% to 75%.

Differentiator: Every estimate shows its evidence, and the model's error rate is published, including when calibration refuses to apply a correction.

TypeScript, React, tRPC, PostgreSQL, MCP API. Private product.

Agentic AI Systems & LLMOps

Agentic OS: Research-to-Publication Agent Pipeline

A model-agnostic agent system that researches topics, drafts and critiques articles, and publishes only after a human approves.

  • LangGraph workflow: Researcher, Writer, Critic, human checkpoint, then Publisher, fed by daily Hacker News, arXiv, and Medium intelligence syncs
  • Model gateway with a LanceDB semantic cache, local MLX and cloud/OpenRouter model routing, and OpenTelemetry trace spans
  • Scheduled GitHub Actions pipeline with Discord approve/reject, cross-machine draft handoff, topic denylists, health monitoring, and 359 tests

Shipped: Articles published to a live blog through the approval loop. Defects found in real production runs were traced to root cause and fixed with regression tests.

View published blog

Serverless Document AI on AWS

Utility Bill Extractor: Textract with Selective Bedrock Fallback

An event-driven AWS pipeline that converts utility-bill PDFs into a stable ten-field JSON schema.

  • S3 and Lambda ingestion, Textract Queries, and SNS/SQS fan-out to a processor Lambda with a dead-letter queue for repeated failures
  • Bedrock structured output is called only for fields that fail the confidence and validation policy. Its answers are merged only after passing the same deterministic Python rules.
  • CloudWatch EMF metrics, raw OCR retention, frozen baselines, and privacy-safe aggregate evaluation

Measured: Development accuracy improved from 66% to 98% across versions, with 92% field accuracy on a held-out set. The held-out result, not the development figure, is the one claimed.

Data Engineering & Signal Detection

Hiring Signal: Daily Hiring-Velocity Intelligence Pipeline

A scheduled pipeline that turns daily job-board snapshots into a time series showing which companies are hiring for which roles, and how fast.

  • 12 sources (Greenhouse, Lever, Ashby, Hacker News and others) across a 396-company watchlist. Each source passed a robots.txt and terms-of-service risk review.
  • Append-only sightings keyed on company and posting URL, which fixed a 48% undercount. A 19-category role taxonomy labels about 72% of 8,918 live listings.
  • Velocity, new-company, and spike alerts, each sent with the SQL that reproduces it, plus immutable per-run archives and rebuild-from-archive recovery

Differentiator: Scraping infrastructure from a personal tool, rebuilt into a data product. The pivot is documented in a written case study.

Python, SQLite, GitHub Actions, Discord alerts. Private repository.

Open-Source Contribution

PrivateGPT: Workbench Chat UX Fixes & UI Test Harness

Contributed to zylon-ai/private-gpt, the widely used open-source API layer for building private AI applications on local models.

  • Fixed streaming auto-scroll, fixed-width chat columns, uncollapsible long messages, and missing copy actions, plus two further bugs found along the way
  • Added a dependency-free Node harness that runs inline UI functions against a fake DOM. Its 20 tests were each confirmed to fail against deliberately broken code.
  • Full project suite passed (2,386 tests). Layout, streaming, and clipboard behavior were measured against a live server.

Full-Stack Product Engineering

ESL Tutoring Pro: Subscription Learning & Digital-Content Platform

A complete WordPress platform for ESL tutors and learners. It organizes educational content, controls subscriber access, and manages digital downloads for a commercial tutoring-resource business.

  • Custom responsive block theme for public, account, lesson-library, subscription, and content-discovery pages
  • Lesson metadata, taxonomies, subscription-product mapping, monthly download quotas, protected downloads, and logging
  • Local development workspace, testing procedures, staging deployment, and ongoing operating workflow

Differentiator: End-to-end ownership of product strategy, UX, engineering, business rules, testing, and deployment.

View live platform

More on GitHub: hybrid-retrieval RAG (BM25, FAISS, rank fusion) and a default-risk MLOps service (MLflow, Docker, CI/CD, Cloud Run).

Explore more on GitHub

SELECTED WORK

Enterprise AI/ML Systems

Large-Scale Data Engineering

Package Network Fingerprint & Temperature Intelligence

Large-scale Azure Databricks/PySpark pipeline that derives package-network fingerprint and temperature intelligence from multi-billion-row operational source data.

  • Native PySpark transformations with partition-aware joins, window operations, column pruning, and selective caching, with no Python UDFs
  • Approximately 4.5 billion source rows producing ~200 million Fingerprint records and ~22 million Temperature outputs with a normalized score from cold 0 to hot 1
  • Row-count reconciliation, duplicate detection, and null/range validation

Scale: Multi-billion-row governed transforms on Azure Databricks without exposing proprietary source tables or identifiers.

Enterprise GenAI

Governed Self-Service Operations Intelligence

Employee-facing AI system that routes natural-language requests between structured-data querying and cited document retrieval under enterprise governance controls.

  • Natural-language-to-SQL with schema grounding, approved patterns, validation, plausibility checks, and restricted execution
  • Azure OpenAI and authenticated enterprise APIs with Azure AI Search hybrid retrieval, semantic chunking, reranking, citations, and answerability thresholds
  • MLflow evaluation for retrieval precision, answer quality, hallucination rate, latency, and failure modes, with human review where operational risk required it

Impact: Governed self-service answers with citation support and fail-closed validation before execution.

Document AI

DOT Document Intelligence & Decision Support

Document intelligence workflow for DOT decision support, combining extraction, reviewer queues, and LLM-assisted evidence validation.

  • AWS Textract, custom neural networks, and Python normalization for images and Word documents, with LLM-assisted evidence validation
  • Reviewer queues with lineage, confidence indicators, and exception handling
  • Source project records report 70% faster processing at approximately 90% extraction accuracy

Reported impact: 60% lower decision time and 35% higher data accuracy.

Applied Machine Learning

Predictive Maintenance Operating Model

Closed-loop predictive maintenance spanning telemetry, model alerts, and technician-facing service recommendations across manufacturing sites.

  • ARIMA, XGBoost/LightGBM, anomaly detection, and computer vision applied to distinct maintenance problems
  • Event-driven alerts with technician-facing service recommendations
  • AWS IoT capability scaled across 50+ manufacturing sites in a closed loop from telemetry and validation through alerting and technician feedback

Source records report: 60% lower equipment downtime and approximately $1 million in additional service revenue.

EXPERIENCE

AI Engineer

FedEx via TEKsystems
  • Contributed to billion-row package-network intelligence pipelines in Azure Databricks and native PySpark, including Fingerprint and Temperature transforms with validation gates.
  • Built and implemented governed natural-language querying and cited document retrieval for employee-facing operations intelligence, with schema grounding, restricted execution, and human review where risk required it.
October 2025 – March 2026

Senior Data Scientist / AI Solutions Architect

HNTB Corporation
  • DOT document intelligence and decision support using AWS Textract, custom neural networks, reviewer queues, and LLM-assisted evidence validation.
  • Reusable ML release and monitoring platform with standardized health monitoring across 15+ models and an approximately 2 TB/day anomaly-detection pipeline.
  • Governed RAG assistant and centralized AI Gateway evaluated on answer quality, citation support, latency, and cost; reported 60% reduction in deployment time.
August 2023 – March 2025

Data Scientist

HCL Global, Capital One engagement
  • Built PySpark ETL on AWS EMR/Hadoop processing billions of rows monthly, with Snowflake, Python, SQL, and Salesforce API integration.
  • Implemented Airflow batch orchestration and Kafka streaming inputs for model-development workflows.
  • Developed XGBoost, LightGBM, and TensorFlow models; transformer semantic-similarity features reducing false positives by 22%.
  • Tracked experiments with MLflow across 200+ recorded model versions and contributed Docker/Kubernetes/SageMaker deployment into an existing serving environment.
July 2022 – June 2023

Data Scientist

Home Depot via Ugam Solutions
  • Retail product-review and merchandising signal analysis with reproducible feature and scoring datasets.
  • A/B-testing frameworks and Hugging Face multilingual review processing across eight languages (~100,000 product reviews per month).
  • BigQuery scoring/reporting automation; reported 30% lower processing time.
February 2022 – June 2022

Data Scientist

Crown Equipment Corporation
  • Connected-equipment analytics foundation for predictive maintenance and service recommendations.
  • AWS IoT program scaled across 50+ manufacturing sites with technician feedback and operational adoption loops.
  • Source records report 60% lower downtime and approximately $1 million in attributed additional service revenue.
August 2019 – January 2022

Earlier Analytics & Applied Data Science

Ugam Solutions · HDFC Bank · PRM Fincon · Evolvalytics
  • Built credit-risk classification models with scikit-learn, improving risk prediction accuracy by 20%.
  • Automated data and reporting workflows, reducing manual effort by 40%.
  • Statistical testing, forecasting, segmentation, and inventory decision support.
2011–2017

SKILLS

GenAI & Agentic Systems

RAG and hybrid retrieval · Semantic chunking and reranking · Citation enforcement and hallucination controls · LangChain · LangGraph · LlamaIndex · Tool/function calling · Human-in-the-loop workflows · Orchestrator/executor patterns

MLOps / LLMOps

MLflow · Docker · Kubernetes · Jenkins · GitHub Actions · Experiment tracking · Model and prompt evaluation · Data/model validation gates · Drift and performance monitoring · Audit logging and rollback planning

Data, Cloud & Platforms

Databricks · Snowflake · AWS EMR/S3/Lambda/SageMaker/Textract · Azure AI Foundry/OpenAI/AI Search/ML/Data Factory · GCP/BigQuery · Kafka and Airflow

ML & Applied Statistics

scikit-learn · XGBoost · LightGBM · TensorFlow · PyTorch · Forecasting · Anomaly detection · NLP · A/B testing · Explainability and bias checks

Core / Primary

Python · SQL · PySpark/Spark · Statistical modeling · Feature engineering · ETL/ELT architecture · Experimentation and model validation

LEADERSHIP & ENGINEERING APPROACH

Technical Leadership

  • Mentored junior data scientists, AI engineers, and interns
  • Supported sprint planning, architecture discussions, and technical standards

Enterprise Translation

  • Converted ambiguous operational problems into measurable systems
  • Presented architecture, risk, and platform trade-offs to technical and executive stakeholders
  • Coordinated data, ML, platform, security, and governance teams

Engineering Principles

  • Reproducibility
  • Explicit validation
  • Observability
  • Failure analysis
  • Security and governance
  • Human review where risk requires it

EDUCATION

Florida State University

Master of Science in Statistical Data Science
August 2017 – May 2019

Sardar Patel University

Master of Science in Statistics
August 2012 – May 2014

CONTACT

Open to discussing AI engineering, machine learning, statistical modeling, data/ML platforms and AI automation work.

vjsojitra8@gmail.com

Email GitHub LinkedIn