AI-powered corporate finance audit system that investigates suspicious payments, duplicate invoices, ghost vendors, policy breaches, and risky transactions — with human approval gates before consequential actions.
FaultAuditAI is a public hackathon project and live AI agent demo focused on controlled, auditable AI workflows for finance teams. Instead of letting an AI agent take risky actions automatically, the system routes findings through approval gates, audit logs, and report generation.
Built as a solo AI engineering project and submitted to a hackathon.
Four pillars — from hands-on engineering to whole-system architecture.
LLM apps, AI agents, RAG systems, FastAPI backends, and production workflows.
OpenAI, Claude, Gemini, Qwen, YandexGPT, APIs, CRMs, databases, dashboards, and automation tools.
Computer vision, dataset workflows, model validation, MLflow, training pipelines, and evaluation.
System design, agent orchestration, human approval gates, audit logs, security, and deployment strategy.
Production-style builds and research across agents, RAG, voice, and vision. NDA-safe summaries.
A repeatable path from a messy workflow to a system you can trust, audit, and operate.
Map the workflow, data sources, constraints, and what success actually looks like — before any code.
Design the agent graph, data flow, retrieval strategy, and guardrails. Decide what stays human-controlled.
FastAPI services, agents, RAG pipelines, and integrations — engineered to be tested and maintained.
Eval sets, MLflow tracking, and validation against real cases so quality is measured, not assumed.
Approval gates, audit logs, and safe rollouts so the system can be trusted in production.
Dockerized deployment, monitoring, documentation, and knowledge transfer to your team.
Real AI products live or die on the surrounding system: routing, retrieval, guardrails, evaluation, approval gates, and audit trails. I design the whole graph — so behaviour stays predictable when it meets real data and real users.
Short technical write-ups on the patterns I use in production.
The engineering behind Callie: a hard real-time audio boundary, three-step barge-in handling, database-enforced idempotency, and schema-validated post-call extraction.
Read →MIT's Recursive Language Models treat huge documents as an environment to query — holding reasoning quality across 10M+ tokens at lower cost and +29% accuracy.
Read →Why reliable enterprise AI comes from architects who design evaluation loops, audit trails, and human oversight — not from prompt users who ship technical debt.
Read →The languages, models, and tools I use to design, build, and ship production AI systems.
I'm a senior full-stack and AI/ML engineer with 14+ years of software engineering experience. I design and build practical AI systems: AI agents, RAG applications, AI SaaS MVPs, voice AI, computer vision pipelines, ML validation workflows, and backend infrastructure. My focus is production-ready AI — systems that can be integrated, tested, deployed, monitored, and improved.
