Tanmay Sule.

Engineers were spending hours manually triaging production alerts, debugging build failures, and searching internal docs. I built the agent framework from scratch and 3 production LLM agents that automate these workflows, reducing time-to-resolution across the engineering organization. Mentored a team of new engineers into AI engineering while establishing org-wide standards.

• Agent Framework — Custom 4-layer architecture on LangGraph/LangChain. Reusable abstractions with built-in observability middleware (per-invocation token tracking), multi-model LLM factory (GPT-5/4o/4o-mini), and centralized tool registry (9 tools).
• Sentry Production Triager — 3-stage pipeline: root cause analysis (GPT-5, 7 tools) → code fix classification → GitHub issue creation with automatic Copilot assignment for code fix. End-to-end AI-to-AI handoff: agent triages the issue, then assigns it to Copilot to generate the fix. Redis distributed locking for concurrent issue prevention.
• Build Failure Triager — Per-failure root cause analysis with cross-failure deduplication. Same root cause across multiple test failures → single GitHub issue, eliminating duplicate engineering effort. Multi-channel Slack reporting.
• Developer Chatbot — Slack-native conversational assistant that answers engineering questions by orchestrating specialized sub-agents. 8-step pipeline with intent classification, dynamic routing, concurrent tool execution, and structured Block Kit responses. Built-in multi-level error recovery.

Full-stack conversational AI platform I designed and built from scratch. The core challenge: answering cross-domain business queries that require joining structured database records with unstructured communications (Slack, Email, Notion) — resolving entity references across data sources and producing evidence-based, source-attributed responses.

• Agent Orchestration — Two-level architecture: L0 orchestrator (9-node LangGraph, 29-field state) handles classification, planning, and evaluation with persistent working memory (entity tracking, pronoun resolution, topic continuity across turns). L1 domain agents (5-node, plan-then-execute) decompose queries into DAGs of typed/SQL/search tasks with concurrent execution. Two-tier human-in-the-loop with crash-safe DynamoDB persistence.
• Data Pipeline — Go ETL with medallion architecture (Bronze → Silver → Gold). 5 event-driven Lambdas, 6 source adapters, self-registering mapper registry normalizing 40+ entity types. Content enrichment prepends source context (channel, sender, date) before embedding — so vectors capture provenance, improving filtered retrieval. Hybrid chunking, dual embedding providers (OpenAI + Bedrock Titan). Multi-layer error tracking with partial-success handling.
• Retrieval Engine — Corrective RAG with hybrid search: parallel pgvector cosine similarity and PostgreSQL full-text search (tsvector, ts_rank_cd weighting) over 50 candidates each, fused via single-pass Reciprocal Rank Fusion SQL (k=60, FULL OUTER JOIN). Cross-source fan-out across all data sources with global re-ranking. LLM-based relevance grading with automatic query reformulation when confidence is low. Separate JSONB metadata filter path for structured queries (channel, sender, date range).
• MCP Tool Layer — 30 tools across 3 Lambda backends: 24 typed ORM handlers (Go, ~20 entity types), 2 SQL tools with read-only guardrails (write blocklist, identifier validation, 500-row cap), 4 search tools (Python). JWT-based tenant authentication (Cognito). Bedrock AgentCore runtime with 13 SSE event types for real-time streaming.
• Infrastructure — 8 ARM64 Graviton Lambdas, Pulumi IaC (Go), per-tenant isolation (PostgreSQL schema-per-tenant, DynamoDB partitioning, dedicated MCP gateway instances). Provider-agnostic LLM factory supporting Anthropic and OpenAI, configurable per tenant.