AI Engineer & Generative AI Researcher | LangGraph · RAG · LLM Evaluation · Multi-Agent Systems · Python · SQL · BI

I make AI prove itself before the business depends on it. 🚀

LangGraph · RAG · GraphRAG · LLM Evaluation · Multi-Agent Systems · Text-to-SQL · Python · SQL · Pinecone · Neo4j · Snowflake · dbt · Power BI · Tableau

• Agentic RAG pipelines with hybrid retrieval, citation validation, and sufficiency guardrails - systems that say "I don't know" instead of hallucinating
• Governed text-to-SQL agents scoped to certified dbt metrics, so the LLM cannot invent numbers that don't exist in the semantic layer
• GraphRAG systems combining Neo4j knowledge graphs with vector retrieval for multi-hop relationship reasoning flat vector search can't do
• Multi-agent evaluation systems that audit AI output before it reaches a human, with schema-validated contracts between agents
• LLM observability stacks scoring faithfulness, cost, latency, and hallucination rates with LangSmith, Langfuse, Ragas, and Grafana

AI/ML: LangGraph · LangChain · Anthropic Claude · RAG · GraphRAG · Pinecone · BM25 · Ragas · DeepEval · LangSmith · Langfuse · Cross-Encoder Reranking
Graph & Data: Neo4j · Cypher · pgvector · DuckDB · Snowflake · dbt · Azure AI Search · ETL · Pandas · NumPy
BI: Power BI · Tableau · Looker Studio · DAX
Infra: FastAPI · FastMCP · Streamlit · Redis · Prometheus · Grafana · Docker · GitHub Actions · Vercel · Railway · Fly.io

The LLM cannot invent a metric that doesn't exist — every answer traces back to a certified dbt model

Business teams wait 3–7 days for analysts to answer questions like "what is 30-day retention for the EU cohort, adjusted for refunds?" The deeper problem is metric drift: "active user" means something different across five dashboards. MetricMind solves both.

• 5-node LangGraph pipeline: Intent Classifier → 3-layer Guardrail → SQL Generator → DuckDB Executor → Response Node, with automatic self-correction on broken SQL
• 3-layer guardrail: PII regex + SQL injection regex + metric allowlist via Claude Haiku — bad queries rejected for $0.0003 vs $0.006 for full pipeline (20x cheaper)
• Governed semantic layer: 6 certified metrics in a JSON catalog; the agent is physically scoped to only those — no hallucinated numbers, no schema drift
• dbt Core: 4 staging models + 4 mart models (DAU, cohort retention, revenue, funnel) with 40 dbt tests catching 150+ dirty rows before any reach a prompt
• 100% eval accuracy on 50-question golden set scored via sqlglot AST comparison · prompt caching on the 3,000-token metric catalog (90% cache hit rate, ~$0.006/query avg)
• Dual anomaly detection: 3-sigma rolling window + Prophet, with HITL commentary approval before publishing
• Full LLMOps stack: LangSmith traces every node · Prometheus scrapes FastAPI every 15s · Grafana dashboard (latency, cost, guardrail rejections) · Tableau Public dashboards
• React + Vite frontend · FastAPI backend on Railway · Vercel deploy · Docker Compose local stack

🔗 Repo · Live Demo

An AI validation finding has to prove itself against real evidence, or it never gets written

Under the SR 26-2 model-risk regime, every GenAI and agentic system a bank runs needs conceptual soundness review, outcomes analysis, and ongoing monitoring. Most tools for this are LLM-as-judge systems, and a judge that hallucinates "faithfulness collapsed to 0.42" is dangerous precisely because it sounds credible. Attestor blocks ungrounded findings at write time.

• LangGraph validation pipeline: Supervisor (materiality-tier routing) → 3 parallel pillar nodes (Conceptual Soundness · Outcomes Analysis · Ongoing Monitoring) → adversarial Challenge node → Attribution Gate
• Deterministic attribution gate: every finding's evidence is resolved against Postgres and its cited metric checked against the stored eval run before persisting. A database lookup, not a second LLM call. Finding.evidence_id is NOT NULL
• Dual-judge scoring: Claude and Llama 3.1 (Groq) score each rubric criterion independently; disagreement beyond threshold escalates to mandatory human review
• Real HITL interrupt/resume: LangGraph interrupt_before on a Postgres checkpointer, so paused runs survive API restarts and are visible across replicas
• Segregation of duties enforced server-side: Keycloak OIDC + RBAC; validators can't validate models they own, and sign-off is MRM-Head-only
• SHA-256 hash-chained audit log · prompt-injection screening · deterministic materiality scorecard · rejected findings feed a self-governance golden set
• Integration tests prove negative controls are rejected: fabricated evidence, cross-model citations, misreported metrics
• Next.js 16 + React 19 · FastAPI · Qdrant · ARQ/Redis · isolated FastMCP server · Logfire + LangSmith · CI with bandit, gitleaks, dependency audit, and a blocking eval-regression gate

🔗 Repo

Reads an invoice image → extracts clean structured data → answers multi-hop supplier questions that vector search can't

Finance teams manually key 50,000+ invoices/month at ~$3.50/invoice. Pure vector RAG can't answer "which suppliers tied to delayed Q3 POs also had quality complaints in the past 18 months?" — because it has no concept of graph structure. LedgerLens solves both halves.

• Claude vision extraction with per-field confidence scoring (0.0–1.0) and automatic human-review routing for low-confidence documents — no OCR pre-processing required
• Neo4j knowledge graph maps Supplier → Invoice → LineItem → PO for relationship reasoning across entities
• LangGraph GraphRAG agent: plan → retrieve → traverse → answer, returning the full Cypher traversal path as an auditable explanation
• 84.2% field extraction accuracy · 99.7% cost reduction vs $3.50/invoice manual baseline (~$0.008/invoice with Claude Sonnet)
• DeepEval/RAGAS eval harness · Langfuse span-level tracing + per-document token cost · FastAPI + React UI · Docker + Fly.io deploy

🔗 Repo

Sales transcript in, cited and grounded live demo out, in minutes instead of days

Solution Engineers spend 5–25 hours a week hand-building demos in a prospect's language. Overture reads the discovery-call transcript, extracts what the prospect needs with every claim traced to an exact quote, and stands up a live Q&A demo that refuses to answer what the source can't support.

• LangGraph extraction pipeline: 4 parallel passes (pains · constraints · requirements · vocabulary) → scope classification → blueprint selection → config fill → validate → persist → embed → share token
• Quote-grounded extraction: every item must carry an exact transcript quote; anything the model can't ground is dropped
• Deterministic where it matters: blueprint picked by a keyword scorer, and the deploy validator has zero LLM imports, proven by a test that greps its own source
• Grounded RAG Q&A over pgvector with expandable citations; verified live refusing a prompt-override attempt to "just figure it out"
• Azure via Terraform: Container Apps · Postgres Flexible Server + pgvector · Key Vault via Managed Identity (zero stored secrets) · Application Insights
• Single container serves React SPA + FastAPI (no CORS surface in prod) · SSE streams live pipeline progress · swappable Claude / Azure OpenAI provider
• 75+ tests · ruff + mypy strict · 50-entry decisions.md of real bugs, e.g. a flaky token-tamper test traced to base64 trailing-byte redundancy (6.6%, measured over 10,000 runs)

🔗 Repo

Every sentence in the SAR narrative is provably grounded in evidence, or refused outright

Transaction monitoring alerts run 90–95% false positives, and a Suspicious Activity Report is a legal filing with FinCEN. A hallucinated sentence isn't a UX bug, it's a false statement to the government. CaseWeave is built around that constraint.

• 9-node LangGraph supervisor: triage (Claude Haiku 4.5) → evidence gathering → narrative (Claude Sonnet 5) → per-sentence guardrail gate
• Frozen EvidenceLedger: every fact gets an ID before drafting, and the narrative model sees only the ledger, never raw rows
• Sentence-level attribution validator with independent judge entailment; below 90% coverage it refuses and returns an evidence-gap report. Caught a real hallucinated regulatory citation in live testing
• Graph-only detection: Neo4j Cypher catches circular fund flows invisible to SQL. Planted-subject recall 91% → 100%
• L0–L4 autonomy ladder per rule; L4 auto-close only with golden-set evidence re-verified at runtime. Same alert: 13 LLM calls / $0.03 at L2 vs 0 / $0.00 at L0
• Redpanda streaming with DLQ · River online anomaly scoring · hybrid pgvector retrieval · DuckDB long-term memory of prior dispositions
• LangSmith + Logfire tracing down to individual guardrail sentence decisions · 5-tool read-only FastMCP server
• 69 tests · CI with ruff, mypy strict, bandit, pip-audit, golden-set gate, and a reproducibility gate · React + Vite + Cytoscape.js review console · synthetic data only

🔗 Repo

• Building production-grade agentic AI evaluation frameworks
• Deepening expertise in LLM observability (LangSmith, Langfuse, Ragas, DeepEval, Grafana)
• Expanding into GraphRAG, multimodal AI, and governed analytics for enterprise use cases
• Targeting AI Engineer / AI Analyst roles at the intersection of AI systems and analytics

• 🎓 MS Information Systems — San Diego State University (GPA: 3.7)
• 🎓 B.Tech Computer Science & Business Systems — DY Patil College of Engineering (GPA: 3.8)

Most people ask "does the AI return an answer?"
I ask "is the answer faithful, grounded, and verifiable — and what happens when it isn't?"

⭐ If my work is useful, feel free to explore the repos!