TPRM Platform with AI & MCP
Publicada el 2026-07-27
Descripción de la oferta
We run a production risk and compliance platform used by large BFSI clients for questionnaire-based assessments, document evidence collection and evidence-backed scoring. We are building an AI layer into this platform across three areas: validating uploaded evidence (document classification, OCR extraction, cross-verification against external registries, tamper and duplicate detection), auto-filling assessment questionnaires from evidence and API data, and auto-reviewing submitted responses (completeness, answer–evidence consistency, contradiction detection, framework-based grading and draft review summaries). All of it operates inside a maker–checker workflow, so confidence scoring and human-in-the-loop design are core to the build, not an afterthought. The skill set we need: hands-on experience integrating LLMs into production applications (OpenAI/Anthropic/open-source models via API), including prompt design, structured outputs, function/tool calling and evaluation of output quality. Document AI experience — OCR pipelines, PDF/image parsing, entity extraction and document classification — since most inputs are certificates, policies and scanned evidence. RAG fundamentals: chunking, embeddings, vector search and citation-grounded answers, so every AI response links back to its source. Familiarity with the Model Context Protocol (MCP) or a strong grasp of tool-calling architectures, as external verification sources will be wired in as governed tools. On the engineering side: Node.js APIs, MongoDB, and comfort working inside an existing Angular product; AWS (S3, EC2) for deployment. Given the domain, the person must think about data security, auditability and PII handling by default — every AI action needs to be logged, scoped and explainable to a bank's audit team. Nice to have: experience with vision models on real-world photo evidence, speech-to-text pipelines, and prior work in fintech, compliance or other regulated environments where "the model said so" is not an acceptable answer.
Skills
Fuente original: freelancer