Skip to content

Christine Straub · Lead Machine Learning Engineer · Remote / United States

Christine Straub · Lead Machine Learning Engineer. Models and agents, measured before they ship.

I build document AI pipelines, reinforcement learning environments that train and grade LLM coding agents, and runtime security for AI agents.

Open to remote Reinforcement Learning, Post-Training, and Agentic AI roles

  • Document AI & Vision-Language Models
  • RL Environments & Agent Evaluation
  • AI Agent Security
  • Computer Vision Data Engines

Christine Straub

I'm a Lead Machine Learning Engineer with 9+ years of shipping production systems in defense, healthcare, legal records and open source. At Sapient Logic I led the team that took two defense AI prototypes to production-ready systems. At Unstructured I merged 145 pull requests into the open-source document parser.

My work falls into three areas. Document AI: OCR, layout models and Qwen3-VL served on vLLM, in cascades that cut per-page extraction cost about 25x while holding accuracy. Reinforcement learning environments: I led 15–20 engineers building the environments frontier labs use to train and evaluate coding agents, including the verifiers, reward signals and anti-cheating checks. AI agent security: LLM-as-a-judge detectors for prompt injection and social engineering in AEGIS, a patent-pending runtime layer at Wider Security.

I care about what happens around the model: how the data is curated, how the output is evaluated, what it costs per page or per rollout, and whether the result is something a person can act on.

Lead Machine Learning Engineer · Remote / United States

Selected outcomes

years building production AI, machine learning, and software systems
9+
merged open-source pull requests to Unstructured (15K+ GitHub stars)
145
lower per-page LLM extraction cost with Qwen3-VL first-line routing
25x
engineers led building reinforcement learning environments for frontier AI labs
15–20
recall on difficult land-record documents in evaluation with a readability-routing stage
48→90%
lower peak memory in Unstructured's open-source hi_res PDF parser (12.9 GB+ to 848 MB)
15x

Four production systems

Each case says what the system is, what I built, and the measured gains.

Reinforcement Learning Environments & Agent Evaluation for LLM Coding Agents

Bespoke Labs

Bespoke Labs builds reinforcement learning environments that frontier AI labs use to train and grade LLM coding agents. I led a team of 15–20 engineers building two of them, including one that drops an agent into a simulated machine learning platform with realistic production failures and checks whether it repairs them.

  • Python
  • Docker
  • FastAPI
  • PostgreSQL
  • XGBoost
  • pytest
  • Claude Code
  • Codex

Outcomes

  • Led machine learning engineering for an agent evaluation platform that runs sandboxed, multi-turn agent sessions across several LLM providers and grades them automatically
  • Built a pass-rate model that rejects 10%+ of weak tasks before human review, so reviewers spend their time on tasks worth keeping
  • Main contributor to a cheat-resistant benchmark: hidden tests, CI promotion gates and locked-down containers, so an agent can only earn the reward by doing the work
  • Built an environment linter as Claude Code and Codex skills that audits every task for ways an agent could game it

Multi-Agent RAG, Model Context Protocol Tooling and LLM Gateways

Duxre · Unstructured · Bespoke Labs

Agent systems in production: a multi-agent assistant for commercial real estate teams at Duxre, Pydantic AI and Model Context Protocol orchestration over document sources at Unstructured, and a multi-provider LLM gateway at Bespoke Labs.

  • Python
  • Model Context Protocol
  • Pydantic AI
  • Azure AI Search
  • GPT-4o
  • RAGAS
  • FastAPI
  • PostgreSQL
  • Claude

Outcomes

  • Built the orchestration layer of Duxre's assistant: planner, executor and presenter agents behind a cost-aware router
  • Gated every prompt and model change in CI on a RAGAS evaluation (faithfulness, answer relevancy, context precision and recall)
  • Enforced role-based access control on retrieval, so each user only gets answers from data their role allows
  • Built Pydantic AI + MCP orchestration with custom MCP servers at Unstructured, and evaluated 10+ vision-language model providers for table and image extraction

Document AI: OCR, Layout Models and Vision-Language Extraction

Medici Land Governance · Unstructured

Extraction of structured fields from land records (deeds, liens, mortgages, court dockets) at Medici Land Governance, and the open-source hi_res PDF and image parser at Unstructured. Both turn scanned and digital documents into structured data for downstream systems.

  • Qwen3-VL
  • vLLM
  • Gemini
  • Claude
  • Instructor
  • Pydantic
  • YOLOX
  • Tesseract
  • PaddleOCR

Outcomes

  • Cut per-page extraction cost about 25x while holding accuracy, by making a self-hosted vision-language model the first-line extractor and escalating only difficult pages to larger models
  • Self-hosted Qwen3-VL on vLLM and tuned batching and memory use for throughput
  • Nearly doubled recall on difficult documents (48% to 90%) in evaluation by adding a readability-routing stage
  • Main contributor to the production document AI platform, including its evaluation service, and wrote its architecture decision records
  • At Unstructured: cut hi_res peak memory on large PDFs from 12.9 GB+ (process killed) to a flat 848 MB, and merged 145 pull requests across 8 repositories

Machine Learning Data Platform and Computer Vision Systems

RIOS Intelligent Machines · Sapient Logic · Collegis Education

Built the data engine behind industrial computer vision models at RIOS: labeling round trips, pipeline orchestration, dataset versioning and leak-free training sets for multi-camera robot video. Earlier computer vision, OCR, speech analytics and cloud data pipelines at Sapient Logic and Collegis Education.

  • Python
  • PyTorch
  • FiftyOne
  • Encord
  • Celery
  • Metaflow
  • Kubernetes
  • ClearML
  • lakeFS
  • Docker

Outcomes

  • Cut the handoff from a flagged event to a labeling task from hours to minutes with a FiftyOne and Encord labeling plugin
  • Moved sequence generation onto parallel Celery workers and rebuilt video-to-dataset creation as a Metaflow flow on Kubernetes
  • Prevented train/test leakage with sequence-level splits and a dataset snapshot before every split
  • Cut new-format data preparation from about a day to minutes with research-format exporters behind one export call
  • Built a ClearML training pipeline and co-built lakeFS dataset versioning

Products I built and shipped

Independent products I designed, built and deployed. Counts come from each codebase.

KINDO home page: headline about knowing a family member's care is handled, beside a care dashboard listing medication, blood pressure, a grocery visit and a cardiology appointment.

KINDO

Live

Family care coordination assistant for adult children and spouses caring for an aging parent or partner, built around hospital discharge and dementia care.

  • Agent registry (orchestrator, pattern scanner, digest, document extraction, coordination planner) behind a provider-agnostic LLM gateway with input guardrails; every AI-drafted action waits for a family member to approve it.
  • Retrieval-augmented chat over pgvector embeddings scoped to each care recipient, plus medication-label capture with a human review step.
  • Supabase backend with 114 migrations, row-level security and 11 edge functions; 11 evaluation suites (6 with LLM judges), and unit, end-to-end, edge-function and database (pgTAP) test suites.
  • Native SwiftUI iOS companion (Swift 6, zero third-party packages): on-device Apple Foundation Models, SwiftData with CloudKit, widgets, Live Activities, Siri and Shortcuts actions (App Intents) and 260 tests.
  • TanStack Start
  • React 19
  • Supabase
  • pgvector
  • Vercel AI SDK
  • SwiftUI
  • Playwright
ShiftMatch home page: headline about scheduling and running care teams, beside a day's visit schedule and an agent relay covering a morning call-out.

Scheduling and operations platform for home care, assisted living, senior living and healthcare staffing agencies.

  • Call-out cascade agent that ranks qualified substitutes by availability, distance, skills, language and overtime, offers the open shift over Twilio SMS, and uses an LLM to classify the replies.
  • Electronic visit verification with GPS geofencing, billing and payroll export, demand forecasting, and a schedule optimizer that only ever produces drafts for a scheduler to publish.
  • Every LLM feature is off by default per agency and stays blocked until the agency acknowledges a business associate agreement; Postgres row-level security isolates tenants.
  • 101 migrations, 19 edge functions, 9 scheduled sweeps, 14 architecture decision records and 200+ test files, with separate staging and production Supabase projects.
  • Next.js 16
  • Supabase
  • PostGIS
  • Twilio
  • OpenAI-compatible LLMs
  • Vitest
  • Playwright
Madre home page: headline about finding help that fits your family, beside an example benefits snapshot with a likely-eligible program, a document checklist and a chat prompt.

Madre

Live

Bilingual (English and Spanish) benefits navigator and community hub for single mothers in the United States.

  • Eligibility computed by deterministic rules, with an AI copilot (Vercel AI SDK) that answers through 7 typed tools over programs, resources, deadlines and applications.
  • Crisis keywords are checked before any model call and routed to crisis lines, so safety does not depend on the model.
  • Semantic program search with pgvector embeddings, 5 background jobs (eligibility checks, reminders, renewal scans, retention, embedding refresh) and a 22-table Postgres schema.
  • 30 unit and 28 end-to-end test files, including automated accessibility checks; built solo in about five weeks.
  • React 19
  • Hono
  • tRPC
  • Drizzle
  • pgvector
  • Inngest
  • OpenAI

Opportunity intelligence and capture management for small government contractors: find, qualify and prepare bids with the evidence attached.

  • SAM.gov ingestion and deterministic qualification scoring with Pursue, Partner, Monitor or Pass recommendations, a human review queue and a pipeline from discovered to won.
  • Capture Studio that assembles statement-of-work drafts and exports Word documents, with multi-tenant roles for owners, capture managers, reviewers and viewers.
  • Python discovery engine behind it: 46 source adapters across federal, state and local procurement, hybrid Postgres full-text and pgvector retrieval fused with reciprocal-rank fusion, and Redis workers.
  • Attachment analysis (deterministic by default, with an optional local model such as Qwen through Ollama) that discards any finding without an exact source quote; digests send only after approval.
  • React 19
  • Hono
  • tRPC
  • Drizzle
  • FastAPI
  • pgvector
  • Redis
Harbor home page: headline about stopping scams before money moves, beside a photo of a parent and adult child with a scam alert card showing a high risk score.

Scam screening for adults 65 and older and the adult children who support them, checking texts, calls and emails before money moves.

  • Two-stage classifier: versioned deterministic rules first, then an LLM verdict service (GPT-4o mini or Claude Haiku); anything below 0.85 confidence is routed to a family member.
  • Continuous-integration evaluation gate over golden, held-out and red-team sets, failing the build below 80% accuracy or above 0.15 expected calibration error.
  • Two Python FastAPI services and a shared library, with a Twilio SMS webhook, web push, a safety-event state machine and consent and data-rights endpoints; 762 Python tests and 4 CI workflows.
  • FastAPI
  • SQLAlchemy
  • Supabase
  • OpenAI
  • Anthropic
  • React 19
  • Twilio

TractWise

MVP

Source-cited review packets for oil and gas land and acquisitions diligence: leases, assignments, title opinions, run sheets and well files.

  • Classifies documents into 9 types and extracts terms with rules, where every field carries its page, verbatim quote and confidence.
  • Cross-document conflict detection and cited keyword question answering that returns "not found" instead of guessing.
  • Reviewer workspace to approve, reject or edit each field, PDF and CSV packet export, and a full audit log.
  • Next.js 14
  • SQLite
  • Clerk
  • pdf.js
  • PDFKit
  • Playwright

Alia Health

Prototype

Referral intake and admission-readiness workspace for skilled nursing facilities that receive referral packets by fax and email.

  • Referral inbox and a three-panel workspace: extracted fields, the source document with each value highlighted where it came from, and an admission checklist.
  • Conflict resolution between documents, a clinical timeline, an audit log, and export that stays locked until a referral is ready for review. Runs on synthetic data.
  • React 19
  • Hono
  • tRPC
  • Drizzle
  • MySQL
  • Playwright

Where I’ve worked

9 roles across defense, healthcare, finance, robotics, legal document workflows, and enterprise AI platforms.

  1. Senior Machine Learning Engineer, AI Security · Wider Security

    Aug 2026 – Present · AI-agent security and containment

    • Develop AEGIS, a patent-pending runtime security layer that supervises AI agents and blocks a compromised or manipulated agent from taking harmful actions
    • Build LLM-based prompt-injection and social-engineering detectors (static pre-checks plus LLM-as-a-judge scoring), validated in shadow mode before they affect enforcement
    • Designed and built the offline evaluation and training harness: reproducible datasets from tamper-evident audit logs, sensitive-data redaction, and red-team go-live benchmarks
    • Python
    • Redis
    • Docker
    • Claude
    • Llama
    • LLM-as-a-judge
  2. Lead Machine Learning Engineer, Document AI · Medici Land Governance

    Jul 2025 – Present · Document intelligence for legal land records (deeds, liens, mortgages, court dockets)

    • Cut per-page LLM extraction cost about 25x while holding accuracy by making a self-hosted vision-language model (Qwen3-VL) the first-line extractor, escalating only difficult pages
    • Self-hosted Qwen3-VL with vLLM on GPUs, tuning batching and memory use for throughput
    • Cut manual review 40% with a deep-learning column-detection pipeline for historical court dockets
    • Gemini
    • Qwen3-VL
    • vLLM
    • Pydantic
    • FastAPI
    • GCP
  3. Lead Machine Learning Engineer, RL Environments & Agent Evaluation · Bespoke Labs

    Jan 2026 – Aug 2026 · Reinforcement learning environments and evaluation infrastructure for coding agents

    • Led a team of 15–20 engineers building RL environments that frontier AI labs use to train and evaluate LLM coding agents
    • Built an AI environment linter (Claude Code and Codex skills) that audits every task for ways an agent could game the reward
    • Led ML engineering for an agent evaluation platform with sandboxed runs across several LLM providers and automated grading
    • Python
    • Docker
    • FastAPI
    • XGBoost
    • verifiers
    • CI
  4. Senior Machine Learning Engineer, ML Data Platform & Computer Vision · RIOS Intelligent Machines

    Jun 2022 – May 2025 · Data engine for industrial computer vision and robotics

    • Built a FiftyOne and Encord labeling plugin that cut the flagged-event-to-labeling handoff from hours to minutes
    • Moved sequence generation onto parallel Celery workers and rebuilt video-to-dataset creation as a Metaflow flow on Kubernetes
    • Wrote sequence-level train/val/test splits so near-duplicate frames never leak across sets; shipped research-format exporters, a ClearML training pipeline and lakeFS dataset versioning
    • Python
    • FiftyOne
    • Encord
    • Celery
    • Metaflow
    • Kubernetes
    • ClearML
    • lakeFS
  5. Senior Machine Learning Engineer, Document AI · Unstructured

    May 2023 – Apr 2025 · Open-source document parsing for RAG and LLM ingestion

    • Merged 145 pull requests across Unstructured-IO repositories (107 to unstructured, 15K+ stars); resolved 300+ bugs and reviewed 500+ PRs
    • Cut hi_res peak memory on large PDFs from 12.9 GB+ (process killed) to a flat 848 MB
    • Fine-tuned YOLOX, YOLO-NAS and Detectron2 on 11,000+ technical PDFs: layout accuracy +10%, missed text -13%
    • Model Context Protocol
    • Pydantic AI
    • OpenAI
    • Claude
    • FastAPI
    • pytest
  6. Lead Software Engineer, AI/ML (Defense & Healthcare) · Sapient Logic

    Jan 2021 – Feb 2024 · Defense AI, healthcare EHR and cybersecurity NLP

    • Technical lead for a team of 5 taking two defense AI prototypes to production-ready machine learning systems
    • Built a BERT/SBERT engine that maps security controls to MITRE ATT&CK: 92% precision, 94% less manual analysis time
    • Architected a HIPAA-compliant EHR on AWS integrating clinical systems over HL7/FHIR
    • Python
    • AWS
    • Django
    • React
    • TensorFlow
    • PyTorch
    • BERT
    • Tesseract4Android
  7. Senior Software Architect, AI Speech Platform · Speechlab AI

    Jun 2022 – Jun 2023 · AI speech-to-speech translation and synthetic dubbing

    • Architected AWS infrastructure for AI dubbing, cutting dubbing costs by up to 80% and turnaround from weeks to hours
    • Integrated speech recognition (ASR), machine translation and text-to-speech (TTS) into the dubbing pipeline
    • Designed and documented the platform API (Swagger/OpenAPI) with AWS Cognito authentication
    • AWS Lambda
    • EventBridge
    • ECS
    • S3
    • Node.js
  8. Senior Machine Learning & Data Engineer · Collegis Education

    Apr 2021 – Apr 2022 · Data and machine learning for higher-education enrollment

    • Reduced student-interaction data latency from hours to minutes with near-real-time GCP pipelines
    • Built the first person-level link between dialer activity and CRM funnel stages with BigQuery and dbt
    • Built a call-recording speech analytics pipeline with Google Cloud Speech-to-Text and BigQuery
    • Python
    • BigQuery
    • dbt
    • Cloud Composer
    • GCP
  9. Software Engineer, Catastrophe Risk Platforms · RMS (Risk Management Solutions), now part of Moody's Analytics

    Sep 2017 – Apr 2021 · Catastrophe risk modeling software for insurers

    • Built client-facing features for RMS(one), RiskLink and RiskBrowser, which insurers use to evaluate multi-billion-dollar portfolios with geospatial risk analysis
    • Java
    • TypeScript
    • Python

Tools I use in production

Grouped by where each tool fits in the systems above.

Reinforcement Learning & Post-Training

  • Supervised Fine-Tuning
  • Reinforcement Learning environments
  • verifiers
  • reward signals
  • benchmark development
  • evaluation harnesses
  • active learning
  • LangSmith

Training & Fine-Tuning

  • DeepSpeed
  • fine-tuning
  • LoRA/PEFT
  • distributed training
  • data curation
  • PyTorch
  • Megatron-LM
  • Hugging Face
  • computer vision
  • NLP
  • TensorFlow

Multi-Agent Workflows

  • multi-agent orchestration
  • agentic workflows
  • AI agents
  • Model Context Protocol
  • tool calling
  • LLM routing
  • context engineering
  • prompt engineering
  • RAG
  • embeddings
  • reranking
  • LangGraph
  • LlamaIndex
  • DSPy

Document Intelligence

  • OCR
  • layout detection
  • structured extraction
  • Vision-Language Model extraction
  • schema validation
  • confidence scoring
  • human-in-the-loop review
  • PDF pipelines
  • Tesseract
  • PaddleOCR

Inference / Optimization

  • vLLM
  • TensorRT-LLM
  • SGLang
  • ONNX
  • quantization
  • A100/H100
  • batching
  • KV-cache
  • latency/throughput benchmarking

MLOps / Cloud

  • Docker
  • Kubernetes
  • CI/CD
  • Metaflow
  • MLflow
  • Weights & Biases
  • observability
  • regression testing
  • GCP
  • AWS
  • FastAPI
  • PostgreSQL

Open source and private work

The Unstructured work is open source. The rest is private client work, summarized here and explained in the case studies above.

Open source: Unstructured

5 repositories linked below

Unstructured's open-source library turns PDFs, images and emails into structured elements for RAG and LLM ingestion. From May 2023 to April 2025 I owned its hi_res PDF and image partitioning path.

  • Rebuilt hi_res: YOLOX detects layout, embedded pdfminer text is merged in, and OCR runs only on regions with no text, so digital PDFs skip most OCR
  • Raised OCR accuracy 15% through Tesseract and PaddleOCR tuning, and added XY-cut sorting for multi-column reading order
  • Cut hi_res peak memory on large PDFs from 12.9 GB+ (process killed) to a flat 848 MB, and fixed encoding and parser failures in email, HTML and XML parsing
  • Maintained the company forks of PaddleOCR and pytesseract, and the Docker image

145 merged pull requests across 8 repositories (107 to unstructured, 15K+ stars), 500+ code reviews and 300+ bugs resolved.

Cheat-resistant benchmark tasks

Main contributor to a reinforcement learning benchmark at Bespoke Labs: hidden tests, CI promotion gates, and a linter that audits every task for ways an agent could game it. Private repository.

Agent evaluation at scale

An agent evaluation platform that runs sandboxed agent sessions across several LLM providers and grades them automatically, with a pass-rate model that filters weak tasks before human review. Private.

Model Context Protocol and multi-agent RAG

Custom MCP servers with Pydantic AI at Unstructured, and the planner/executor/presenter orchestration behind Duxre's assistant, gated in CI on a RAGAS evaluation.

Vision-language extraction evaluation

A hand-labeled gold set and an evaluation service that gates every model release at Medici Land Governance, which made a 25x cheaper first-line model safe to ship.

Computer vision data engine

Labeling round trips, Celery and Metaflow pipelines, sequence-level splits and lakeFS versioning for multi-camera robot video at RIOS.

Résumé

Full employment history and credentials, one page.

Download Resume

Every company and project

34 companies, 64 shipped projects, grouped by domain. Company descriptions come from public sources; engagements without a public client are listed without a company name.

Reinforcement learning & agent evaluation

2 engagements

Applied AI research lab that builds the reinforcement learning environments and training data frontier AI labs use to train and evaluate AI agents.

Led a team of 15–20 engineers building two reinforcement learning environments that frontier AI labs use to train and evaluate LLM coding agentsLed machine learning engineering for an agent evaluation platform that grades sandboxed agent sessions across several LLM providers

2 projects

Agent evaluation platform

Led machine learning engineering for a production agent evaluation system that delivers reinforcement learning environment tasks, with verifiers and reward signals, for multi-turn tool-calling agents. Built a pass-rate model that rejects 10%+ of weak tasks before human review, and a gateway in front of several LLM providers that handles keys, budgets and cost tracking.

  • Python
  • FastAPI
  • PostgreSQL
  • XGBoost
  • Docker
  • Cloud Run

Cheat-resistant machine learning platform benchmark

Coding agents repair realistic failures inside a simulated production machine learning platform. Main contributor: CI promotion gates, hidden tests and locked-down containers so an agent can only earn the reward by doing the work, plus an environment linter that audits every task for ways an agent could game it.

  • Python
  • Docker
  • pytest
  • GitHub Actions
  • Claude Code
  • Codex

AfterQuery

Project

Frontier coding tasks: PyTorch systems

Authors difficult, reproducible PyTorch systems tasks for evaluating autonomous frontier coding agents, each shipped with a complete Harbor harness and a verified golden solution, built on vLLM, SGLang, TensorRT-LLM, DeepSpeed and Megatron-LM.

  • PyTorch
  • vLLM
  • SGLang
  • TensorRT-LLM
  • DeepSpeed
  • Megatron-LM

Document intelligence & OCR

3 engagements

Builds open-source and commercial document-ingestion tooling that turns PDFs, Word files, and other unstructured documents into structured data for LLM pipelines.

Cut hi_res peak memory on large PDFs from 12.9 GB+ (process killed) to a flat 848 MB145 merged pull requests across 8 Unstructured-IO repositories

Project

hi_res PDF and image partitioning

Owned the path that turns scanned and digital PDFs into ordered elements for RAG. Split layout detection (YOLOX) from text handling so embedded pdfminer text merges first and OCR runs only on missing regions, then added XY-cut ordering for multi-column pages. OCR accuracy rose 15%.

  • Python
  • YOLOX
  • Tesseract
  • PaddleOCR
  • pdfminer
  • Docker

Medici Land Governance

Private company

Turns scanned public land records (deeds, liens, mortgages and court dockets) into structured, searchable data with a document-AI platform.

Nearly doubled recall on difficult documents (48% to 90%) in evaluation with a readability-routing stageMain contributor to the production document AI platform, including its evaluation service

Project

Legal document extraction with difficulty-aware routing

Extracts structured fields from deeds, liens, mortgages and court dockets. A self-hosted vision-language model handles routine pages and only difficult pages escalate to larger hosted models, which cut per-page extraction cost about 25x while holding accuracy on a hand-labeled gold set.

  • Qwen3-VL
  • vLLM
  • Gemini
  • Claude
  • Pydantic

AI Customer Assistance

Project

Building-defect OCR and NLP

OCR and NLP for residential building-defect communication: extracts text from specification PDFs and classifies it into assistance workflows.

  • Python
  • OCR
  • embeddings
  • document parsing

Generative AI, LLMs & agents

11 engagements

Duxre

Startup

Builds an AI-powered deal platform for commercial real estate brokers, combining listing marketing, CRM, email campaigns, and verified AI-readable deal briefs.

Built the orchestration layer of the platform's AI assistantGated prompt and model changes in CI/CD on a RAGAS evaluation

Project

Multi-agent assistant for commercial real estate

Planner, executor and presenter agents behind a cost-aware router answer questions from a firm's own documents and from public market data. Role-based access control limits each user to data their role allows, and table-aware chunking keeps rent rolls and financial statements intact.

  • Python 3.12
  • FastAPI
  • Azure AI Search
  • Azure OpenAI
  • LangChain
  • RAGAS
  • pgvector

Duarte

Private company

Provides presentation design, brand storytelling, and communication training that helps companies craft and deliver persuasive presentations.

Project

Agentic slide generation

Pipeline that renders uploaded decks to images, has a vision model describe what each slide is doing, and generates new slides from that description, with output validation and tracing.

  • FastAPI
  • Claude 3.5 Sonnet
  • Guardrails AI
  • AgentOps
  • python-pptx
  • Pydantic 2

Builds an AI platform that automates social-media content creation and scheduling for small businesses and marketing agencies.

Project

AI social media content generation platform

Generates social posts in each small business's brand voice. GPT-3 writes the posts, and a brand-analysis engine reads the company's website to extract tone, messaging and style markers. The full-stack app runs on Next.js and Supabase with PostgreSQL.

  • GPT-3
  • Next.js
  • Supabase
  • PostgreSQL
  • Vercel

Education and certifications

A UC Berkeley degree, plus specializations from Stanford, DeepLearning.AI, IBM, AWS and Google.

Education

  • B.A., Computer Science and Cognitive Science (Dec 2017)

    UC Berkeley

Certifications

  • Machine Learning SpecializationStanford University
  • Deep Learning SpecializationDeepLearning.AI
  • IBM Data Science SpecializationIBM
  • AWS Cloud Practitioner EssentialsAWS
  • Certified Scrum Product Owner (CSPO)Scrum Alliance
  • Google Business Intelligence SpecializationGoogle
  • Google Data Analytics SpecializationGoogle

Frequently Asked Questions

What roles is Christine Straub open to?

Christine Straub is open to remote Reinforcement Learning, Post-Training, and Agentic AI roles.

What is Christine Straub's core expertise?

Christine Straub works in three areas: AI agent security (runtime containment and prompt-injection detection at Wider Security), reinforcement learning environments for LLM coding agents (verifiers, reward signals, cheat-resistant tasks), and document AI (OCR, layout models and vision-language extraction at Unstructured and Medici Land Governance). Before that, she built computer vision data engines at RIOS (labeling round trips, dataset versioning and leak-free training splits).

Where has Christine Straub shipped production AI?

Christine Straub has shipped production AI in defense (two defense AI prototypes taken to production-ready systems at Sapient Logic, and a real-time network health dashboard for a U.S. Navy program), healthcare (a HIPAA-compliant electronic health record on AWS), and open source (145 merged pull requests to Unstructured, 15K+ GitHub stars).

What has Christine Straub built with vision-language models?

At Medici Land Governance, Christine Straub made a self-hosted vision-language model (Qwen3-VL on vLLM) the first-line extractor for land records, cutting per-page extraction cost about 25x while holding accuracy, with only difficult pages escalated to larger models.

Has Christine Straub led engineering teams?

Yes. Christine Straub led a team of 15–20 engineers building reinforcement learning environments at Bespoke Labs, a 5-engineer defense AI team at Sapient Logic, and a 4-engineer document extraction team at Medici Land Governance.

Where does Christine Straub work now?

Christine Straub works at Wider Security on AEGIS, a patent-pending runtime security layer that supervises AI agents and blocks a compromised or manipulated agent from taking harmful actions. She builds its prompt-injection and social-engineering detectors (LLM-as-a-judge, validated in shadow mode) and the evaluation harness that benchmarks them against legitimate and red-team agent sessions.

What is Christine Straub's education and background?

Christine Straub holds a B.A. in Computer Science and Cognitive Science from UC Berkeley (2017) and has 9+ years building production machine learning and software systems, starting as a software engineer on catastrophe risk platforms at RMS (now Moody's Analytics).

Contact

If you’re hiring for reinforcement learning environments, agent evaluation, document AI or security for AI agents, email me.

Location
Remote / United States