Nupur

Part of a productOS platform, built for brownfield codebases where a product's features live only in thousands of undocumented functions and tribal knowledge. Glue turns a codebase into a queryable product graph and runs agents on top to auto-document features, answer questions in plain English, and triage bugs to a root cause, owner, and fix plan.

• ✦Deterministic, language-specific indexers (Java, TypeScript, Python, Go) compile a codebase into a queryable product graph: functions as nodes, calls as edges, plus routes, API calls and dependencies
• ✦Louvain community detection clusters the graph into coherent features, replacing the naive whole-graph-to-LLM approach and giving the documentation agent a structured base to work from
• ✦A multi-agent suite on the Claude Agent SDK: a documentation agent (Claude skills), a Q&A agent reachable in Slack, and a triage agent that runs RCA from the graph, production logs and commit history, names the code-owner, and writes a fix plan
• ✦Built the orchestration harness and an agent tool registry (tool-call logging, retries, rate limits, pluggable skills), with integrations across ClickUp, Jira, Linear, GitHub, Slack, Sentry and Search Console feeding the agents their context
• ✦An MCP server exposes the product graph and generated dev plans natively in the IDE (Claude Code, Cursor, Cline)

ticket → fix plan · RCA · owner · dev plan

Leucine's flagship AI bet: a “Lovable for Pharma,” built before Lovable existed. A no-code builder that turns each manufacturer's plain-English SOPs into fully auditable, GMP-compliant apps, replacing bespoke per-client engineering with self-serve configuration, with no prior art to copy and only the weak, hallucination-prone models of the time to build on.

• ✦Built the full LLM orchestration platform from scratch in the pre-agentic era: input/output guardrails, a multi-model gateway (OpenAI, Anthropic, Gemini, Mistral) with fallback routing, rate limiting, and request/cost/error logging
• ✦Made generation deterministic and self-healing: the LLM is constrained to fixed schemas, its SQL validated in a sandbox against synthetic data and auto-repaired from error logs, turning weak, small-context models into production-grade config
• ✦Palantir-inspired ontology-and-actions model: a workflow agent generates per-tenant ontologies, a form/action engine emits ordered side-effects, with multi-tenant isolation via per-tenant schemas
• ✦RAG over uploaded SOPs pulls the finer details, formulas and step constraints, into the generated forms
• ✦GMP compliant by construction: immutable audit trails, e-signatures via re-authentication, strict role-based access

$500K · Piramal · Revlon deployed

A co-investigator for GMP deviation handling at Leucine. An orchestration agent that runs a batch investigation through phases and pinpoints the root cause from the company's own documents, with evidence and confidence.

• ✦Phase-driven investigation agent: classify the deviation, surface similar historical RCAs, then root-cause, advancing only on QA confirmation
• ✦RCA agent grounds findings in indexed SOPs/MBRs via RAG, applying 5-Why and fishbone, presented with evidence and confidence
• ✦Layout-aware ingestion: IBM Docling OCR preserves section hierarchy and tables, section-aware chunking with metadata
• ✦OpenAI 1536-dim embeddings in Postgres/pgvector (no dedicated vector DB), apt for pharma's data-residency constraints and scale

months → hours · RCA, co-investigated with QA