Kaushal Prajapati
Staff AI/ML Engineer · researcher · writer
I build the system layer between models and products: agent runtimes, evaluation gates, retrieval, memory, model routing, tool authorization, and developer platforms.
Research depth.
Platform instincts.
Operator judgment.
My work sits in the layer most AI demos skip: the infrastructure that makes model behavior reliable, observable, secure, and usable by product teams.
I care about leverage. A good platform decision makes every product team faster; a good explanation lets other engineers borrow the lesson without paying the same failure cost. That is why I write here.
- 6+ products · 99.9% success
HighLevel
Staff Software Engineer, AI PlatformDallas, TX · remoteArchitected the unified AI runtime and agent harness behind six AI product surfaces. The platform carries 15M+ conversations and 800B+ tokens per quarter. Also built a headless developer platform with an MCP server, SDK, CLI, and plugin system exposing 600+ operations.
- enterprise compliance AI
Zania AI
Applied AI Engineer, Compliance AgentsSan Francisco, CA · remoteBuilt enterprise compliance and cybersecurity automation using autonomous agents, retrieval-augmented generation, document question answering, and scalable LLM job orchestration.
- product + platform
Intellect AI · Intellect Design Arena
Lead Machine Learning EngineerBengaluru, India · remoteTook Document AI, underwriting, graph intelligence, and information-extraction systems from model architecture through multi-client commercialization.
- 7 production systems
Eagle Labs · First American India
Research & Development EngineerBengaluru, IndiaBuilt production systems across document intake, understanding, verification, property risk, data security, and customer intelligence for the title and escrow lifecycle.
Agents, retrieval, and document intelligence.
Enterprise AI systems that connect model quality to reliability, operating cost, customer retention, and commercial growth.
ZA::01Applied AI Engineer, Compliance AgentsZania AI · San Francisco, CA · remote · Nov 2024 – Jun 202586% → 97% F1@10 · days → hours · largest-client retention
Enterprise compliance and cybersecurity automation built on autonomous agents, RAG, and document question answering.
- Built autonomous agents and intelligent bots that automated SOC 2, NIST, HIPAA, third-party risk assessment (TPRA), and security-questionnaire workflows using LLMs, RAG, and document question answering.
- Raised evidence-retrieval F1@10 from 86% to 97%; the improvement directly prevented churn of Zania's largest enterprise client.
- Built post-training and evaluation release gates with synthetic control-evidence pairs, hard negatives, rejection sampling, and held-out audit QA to measure faithfulness, citation accuracy, hallucination rate, and control coverage.
- Fine-tuned Qwen2.5-7B, Llama-3.1-8B, Phi-3 Mini, and Gemma-2-9B using LoRA/QLoRA, PEFT, and 4-bit quantization, then deployed cost-aware routing between open small language models and frontier LLM APIs.
- Designed and deployed asynchronous LLM job orchestration with Azure Service Bus and Redis, including global rate limiting for high-concurrency workloads.
- Developed a cybersecurity-documentation multi-agent system for Netflix that reduced delivery SLA from days to hours.
IA::02Lead Machine Learning EngineerIntellect AI · Intellect Design Arena · Bengaluru, India · remote · Sep 2021 – Nov 20246M+ PDFs · 23+ → 1 model · 4× sales growth
Commercialized Underwriting CoPilot and unified the ingestion, extraction, graph, routing, and human-in-the-loop systems behind enterprise document intelligence.
- Commercialized Underwriting CoPilot, a multi-agent RAG system that gave underwriters contextual document understanding and dynamically generated risk summaries.
- Engineered layout-aware chunking and document-ingestion microservices that processed more than 6 million PDFs while cutting AWS Textract costs by 70%.
- Fine-tuned 7B-parameter LLMs on domain-specific corpora, improving zero-shot generalization F1 on unseen document types from 73% to 79%.
- Developed an end-to-end NLP pipeline for meeting-transcript analysis that populated Neo4j knowledge graphs with entities, speaker context, relationships, and topics.
- Added natural-language-to-Cypher conversion so users could run analytical graph workflows without writing Cypher queries.
- Implemented an intent-aware query router that selected Boolean, full-text, vector, or LLM pipelines, avoiding unnecessary high-capability model and vector-database calls.
- Applied GraphRAG to title-chain and claims graphs to detect fraud patterns and suspicious relationships.
- Built a unified, layout-aware information-extraction model that replaced 23+ legacy ML pipelines, reducing deployment complexity and technical debt.
- Productized the intelligent document processing service, enabling 4× sales growth and onboarding six new enterprise clients.
- Developed a human-in-the-loop data-labeling system that reduced annotation and model-retraining cycles by 60%.
- Created modular OCR pipelines with document-type-aware routing to the most suitable OCR backend.
Applied ML across the title lifecycle.
Seven production systems spanning document intake, understanding, verification, risk, privacy, and customer intelligence.
FA::01Customized OCR PipelineA text-detection and recognition pipeline for scanned PDFs.2% CER on VINs · GEM of the Year
- Developed a reusable text-detection and recognition pipeline and library, reaching a 2% character error rate (CER) on vehicle identification numbers.
- Benchmarked the pipeline against Tesseract, PaddleOCR, and AWS Textract.
- Built a preprocessing API that analyzed each document and selected the most suitable OCR engine.
- Removed AWS Textract dependencies from multiple services, reducing infrastructure costs and the price of downstream service offerings.
- Improved the performance of document-classification and information-extraction services across multiple applications.
FA::02Financial Document Understanding PlatformA platform for identifying and extracting information from financial and legal documents.2M+ documents · 1,400+ categories · 70% faster SLA · 34% lower costs
- Developed a large-scale data pipeline to classify more than 2 million documents across 1,400+ categories.
- Worked on layout analysis, table detection and extraction, form understanding, and free-text clustering.
- Built information-extraction workflows using BERT for named-entity recognition in unstructured documents and LayoutLM for structured documents.
- Reduced single-document processing SLA by 70% and helped reduce extraneous escrow staffing and service costs by 34%.
- Received the Best Employee Quarterly award for the platform's impact.
FA::03Exploring Variations in Large Document CorpusA pipeline for finding distinct layout variations in large document corpora.layout discovery · corpus-scale clustering
- Experimented with visual and structural representations to identify the most effective embeddings for document-layout variation.
- Researched and evaluated clustering algorithms for grouping unique layout families at corpus scale.
FA::04E-Signing & NotarizationA signature and notary detection, extraction, and verification toolkit.64% → 87% precision · 70% faster SLA
- Designed and implemented a modular, extensible API for signature and notary detection.
- Added a false-positive removal layer that increased model precision from 64% to 87%.
- Developed an algorithm to validate whether notary seals comply with US state notarization laws.
- Built a data-extraction pipeline for notary seals and integrated the toolkit with DocuSign and First American's document-processing platform.
- Reduced signature and notary verification SLA by 70%, helping lower escrow-service costs and increase document-recording revenue.
FA::05Property Risk ProfilingAn AI system for estimating the risk involved in insuring a property.73% → 81% performance · 19% fewer claims
- Developed data pipelines to automate ETL across multiple property and claims data sources.
- Engineered complex geospatial data to discover features influencing a property's risk profile.
- Built a regression model, a unified dashboard, and an automated report-delivery system for risk profiling.
- Integrated Google Maps with customized GIS layers to visualize property boundaries.
- Integrated the First American Claims API, improving model performance from 73% to 81% and reducing title-insurance claims by 19%.
- Received the Best Employee Quarterly award for the project's impact.
FA::06Data Security InspectorA configurable system for identifying and redacting NPI and PII.configurable rules · profile-aware redaction
- Developed an automated, scalable, and resilient pipeline to replace fragmented ETL processes.
- Enabled end users to define NPI and PII detection rules through a user interface.
- Added profile-aware redaction so sensitive data could be protected according to user access and policy.
FA::07Customer SegmentationAn ML system for segmenting First American customers using business-defined signals.automated insights · stakeholder-ready reports
- Built an ETL pipeline to query the First American Transaction System.
- Evaluated clustering algorithms and created visualizations that made customer segments actionable for stakeholders.
- Automated outlier detection using thresholds and criteria supplied through a user interface.
- Created processes to mine patterns and generate report-ready, business-driving insights in natural language.
Tools are choices, not identity.
I choose for reliability, operability, and team velocity.
Working on the hard part of AI?
I’m open to thoughtful conversations about AI platforms, agent infrastructure, and Staff/Principal engineering work.