Jerusan Jegatheeswaran
I am a full-stack engineer specializing in AI-integrated architecture. I go beyond basic wrappers to build deterministic systems around non-deterministic models. My work focuses on orchestrating LLMs and VLMs for unstructured data extraction, engineering multimodal RAG pipelines with strict citation grounding, and designing human-in-the-loop review interfaces. I own features end-to-end—bridging the gap between experimental AI and production readiness—and have shipped everything from high-performance trade execution engines to SOC 2 automation platforms.
Experience
ComplianceOS ↗
Founder // Minneapolis, MN- Built a full-stack SOC 2 automation platform (React + Python) with real-time compliance dashboards and automated control evaluation pipelines, cutting audit prep time from days to hours.
- Shipped Qwen3.5 into production to parse natural language and classify unstructured ticketing data, orchestrating automated evidence collection across identity providers.
- Designed the backend pipeline to integrate across cloud identity providers, significantly reducing manual audit workflows end-to-end without human intervention.
Midpoint Labs
Software Engineer // SF- Architected a high-performance, event-driven Go backend processing 50K+ daily events supporting real-time blockchain request tracking.
- Engineered a custom blockchain listener strategy that eliminated costly third-party API dependencies, reducing external service requests by 80%.
- Designed full-stack observability tooling including user action telemetry and system health monitoring to resolve frontend friction points and optimize system reliability.
Diffuse Funds
Software Engineer // Chicago- Engineered a high-reliability cryptocurrency trade execution engine, optimizing transaction handlers to guarantee trade placement in the earliest available block.
- Built a TypeScript metadata API aggregating smart contract ABIs across 20+ DeFi protocols, cutting new protocol integration time by 80%.
- Automated trade auditing and monthly reconciliation in Python, compressing multi-day manual workflows to ~10 minutes while improving compliance accuracy.
KLDiscovery
Software Engineer // MN- Led a 7-person engineering team to completely overhaul the platform's permission system, delivering enterprise-grade security that directly unlocked large law firm clients.
- Built automated case setup pipelines and a secure cloud storage integration UI, eliminating manual onboarding queues and accelerating customer time-to-value.
- Engineered a multithreaded analytics engine processing 100K+ monthly records to automatically rank and surface reviewer productivity.
Featured Work
AI-Integrated Full-Stack Architecture
UnderLoupe
Answers you can inspect.
A multimodal reasoning workbench that turns an equipment manual into something a technician can query at the machine, with answers carrying figure crops, highlighted diagram callouts and step-by-step wizards. Rebuilt from a hand-authored prototype into a document-agnostic pipeline where one command turns a new document into a retrieval pack, and every claim carries a ref validated against that pack before the stream closes, so an unverifiable citation gets surfaced rather than hidden.
[ live_demo ]DataMill
Extraction you can correct.
A local review terminal for the step RAG pipelines usually run blind: PDF layout parsing. Operators correct bounding boxes on the page while a live pane recompiles the exact markdown that will ship, an OpenCV scanner proposes regions no parser caught, and VLM transcription stays gated behind an explicit click, so no page costs money until someone asks for it. Exports the RAG bundle UnderLoupe answers from.
Dex
Built a voice-based AI interviewer for multi-stage technical assessments with real-time LLM feedback on problem-solving and code quality.
[ source_code ]Selected Projects
Reality Layer
An interactive introspective platform integrating Sonar APIs to process natural language inputs, deliver immediate fact-checking, and generate a visual 'Belief Constellation' mapping the interconnections of user thoughts.
ASVec
Engineered an HNSW approximate nearest neighbor search engine in C++ with SIMD vectorization and concurrent graph construction.
[ source_code ]Tech Stack
Contact
Always open to discussing complex systems, AI architecture, or new engineering roles. Reach me at jerusan@meridianworks.io.