Vishal Murugan
Artificial Intelligence Engineer
I build systems that grade their own reliability - LLM eval harnesses, cost-aware model routers, and drift monitors designed to catch the failures a text-only pass/fail check would miss.
Experience
AI Engineer Intern
Zetheta Algorithms Private Limited
Jan 2026 - Jun 2026
Remote
Multi-Agent Compliance Monitoring System
- ▹Architected a multi-agent compliance monitoring system, coordinating 4 specialized agents (transaction surveillance, NLP communication scanning, regulatory tracking, report generation) through an event-driven pub/sub message bus with a shared alert schema for zero-coupling inter-agent communication
- ▹Designed and implemented a 10-rule real-time transaction surveillance engine detecting AML patterns including structuring, layering, velocity anomalies, and mule-network fan-out, using statistical baselining (z-score deviation) for entity-specific anomaly detection
- ▹Built a hybrid NLP communication scanner combining regex/keyword rules with a TF-IDF cosine-similarity semantic layer, enabling detection of paraphrased compliance violations with zero exact keyword overlap - validated against adversarial test cases
- ▹Engineered a cross-agent correlation and escalation framework that auto-links alerts across agents within a 60-minute window and elevates severity on multi-signal risk, reducing manual triage load by surfacing compound-risk entities first
- ▹Integrated live regulatory data ingestion from SEC/FINRA public RSS feeds with automatic keyword-based relevance-matching against internal detection rules, backed by a persistent SQLite enforcement-action database
- ▹Achieved 100% pass rate across 22 automated validation scenarios (exceeding the 20-scenario spec) covering true positives, true negatives, and edge cases; shipped an interactive Streamlit dashboard with live analytics (escalation rate, false-positive tracking by rule, entity risk scoring)
Projects
A text-to-SQL system that converts natural-language questions into SQL - but with an added layer that detects when a question is genuinely ambiguous and asks a clarifying question instead of guessing.
It's not just "I called an LLM to write SQL" - it demonstrates three things hiring managers actually care about: (1) structured output discipline (Pydantic schemas, not regex-parsed text), (2) an eval-first mindset (a labeled test set + precision/recall scoring, not eyeballing outputs), and (3) the ability to diagnose why a model fails and fix it with targeted iteration rather than blind prompt tweaking.
A retrieval-augmented generation (RAG) system built over an internal engineering knowledge base - but the actual point isn't the RAG pipeline itself, it's the evaluation harness that scientifically compares different ways of building that pipeline. Specifically, it isolates two design choices - how you chunk documents, and how you retrieve relevant chunks - and measures the impact of each combination independently.
An MCP (Model Context Protocol) server that exposes GitHub operations — listing issues, reading PRs, searching code, creating issues, closing issues, merging PRs — as tools an LLM can call. Alongside it, an evaluation harness that tests whether the LLM actually uses those tools correctly and, more importantly, safely.
A benchmarking suite that compares locally-run LLMs (via Ollama) across model size tiers - Llama 3.2 1B, 3B, and 8B - measuring memory footprint, generation latency/throughput, response quality (via LLM-as-judge), and structured-output reliability, all under identical test conditions.
A pipeline that takes a code diff (a pull request), automatically parses it, breaks it into manageable pieces, has an LLM review each piece for real engineering issues (security vulnerabilities, missing error handling, resource leaks, logic bugs, missing test coverage), and produces a structured, line-precise report. It's evaluated not just by "does it run" but by measuring recall (does it actually catch bugs?) and precision (does it stay quiet on correct code?) against a labeled set of real and injected bugs.