About Me
Agentic AI Engineer building, evaluating, securing, and deploying production-grade LLM agents, Enterprise Decision OS platforms, and Agentic RAG systems with Python, LangGraph, FastAPI, and FastMCP. Shipped to real users…
Agentic AI Engineer building, evaluating, securing, and deploying production-grade LLM agents, Enterprise Decision OS platforms, and Agentic RAG systems with Python, LangGraph, FastAPI, and FastMCP. Shipped to real users: an LLM-powered cricket bowling planner used by Japan’s national squad for Asian Games 2026 preparation, ALBA ONE (an 11-phase multi-tenant Decision OS with 22 MCP servers), and customer-facing RAG pipelines. Experienced in hybrid retrieval, DeepEval/LangSmith evaluation, 7-layer AI guardrails, red-teaming, and Kubernetes LLMOps.
Experience
Agentic AI Intern
<p>Engineered an agentic AI bowling machine with an LLM planner as its decision-making brain; built a player knowledge base for Japan’s 15-player national squad from 2 years of ball-by-ball match data, using exploratory data analysis to derive batting strengths, weaknesses, dismissal patterns, and performance under pressure.</p><p> Built tool-augmented planning (database, knowledge base, hybrid retriever, reranker, Text-to-SQL): when a batsman is selected, the agent retrieves and reranks player-specific context, plans each delivery, observes the outcome, and adapts the next ball using thread-scoped short-term memory consolidated into persistent per-batsman long-term memory.</p><p> Ran offline evals with DeepEval: achieved 0.95 task completion, 1.00 step efficiency, 0.80 plan quality, 1.00 plan adherence, 0.87 argument correctness, and 1.00 tool correctness, while retrieval scored 0.99 contextual recall and 0.91 contextual precision.</p><p> Validated production readiness with P95 latency
PROJECTS
ALBA ONE (Enterprise Decision OS)
<p> Architected a production multi-tenant Decision OS evolving a management Digital Twin prototype into a 6-stage LangGraph state</p><p>machine across 22 FastMCP servers (44 typed tools) and a bilingual (EN/JP) Next.js 15 cockpit; built hybrid retrieval</p><p>(PostgreSQL pgvector HNSW + BM25/PGroonga + RRF k = 60 + cross-encoder reranking) with SHA-256 chunk lineage.</p><p>• Engineered claim-level citation verification and an Expected Value of Sample Information (EVSI) decision judge; resolved</p><p>noisy-chunk low contextual relevance via semantic chunking to achieve 0.94 NDCG@10, 0.96 Recall@20, 0.91 MRR, 0.95 faithfulness,</p><p>100% citation verification, and 0.93 tool-call accuracy across 219 automated pytest/DeepEval gates.</p><p>• Enforced 7-layer defense-in-depth guardrails (deterministic regex/Unicode NFKC + semantic injection classifier, APPI/GDPR PII</p><p>redaction, PostgreSQL RLS tenant isolation, MCP HMAC capability tokens, groundedness validator, RBAC HITL gates, and Redis</p><p>token-budget breakers) alongside RS256/ES256 OIDC JWT auth (30s JWKS anti-amplification) and fail-closed startup checks.</p><p>• Shipped non-root (UID 10001) multi-stage Docker images, Helm/Argo Rollouts canary pipelines (auto-rollback at >1% 5xx or P95</p><p>>8s), and OpenTelemetry/Prometheus observability; built an 11-family automated red-team harness (0% ASR on 500+</p><p>adversarial prompts, Wilson 95% CI drift gates, and