context infrastructure, coding-agent systems, and scientific software.

16 · Kolkata, India · TypeScript / Python · currently building Loreflow

I’m mostly interested in the layer around increasingly capable models: context, retrieval, state, provenance, rollback, safety, and the systems that make agents useful outside a demo.

The company brain for fast-moving startups.

Company knowledge gets fragmented across docs, conversations, decisions, meetings, and tools. Loreflow is the context layer that turns that mess into something humans and AI can actually reason over.

connectors → normalized entities → knowledge graph → semantic/vector retrieval → AI reasoning → source citations

Current product stack includes TypeScript, React 19, TanStack Start, Bun/Hono, Clerk Organizations, Supabase/PostgreSQL, and Drizzle. Loreflow is supported by Microsoft for Startups and OpenAI for Startups.

Closed-source scientific software for reaction-constrained molecular discovery. The system keeps proposed routes, provenance, uncertainty, disagreements, failures, and evidence attached to candidates instead of treating synthesizability as an afterthought.

Customer discovery with drug-discovery operators pushed the product toward a route-and-evidence layer for expert review rather than another molecule generator.

A TypeScript SDK for local coding-agent checkpoint / fork / reconcile / rollback loops. Dirty-set restoration makes rollback scale with what the agent changed instead of the whole repository.

npm i hyperion-delta

A Claude Code-native safety harness that intercepts risky tool calls, red-teams repositories against destructive / secret-leaking agent behavior, and generates enforceable hooks + CI/SARIF reports.

pip install butterfence

Structure-aware semantic chunking for RAG pipelines: heading/list/table/code-block preservation, embedding-based boundaries, deduplication, noise removal, and metadata-rich output.

Co-built a local-first AI egress gateway for source code with local repository indexing/search, context-pack generation, privacy-preserving routing, tamper-evident audit trails, and a VS Code/Cursor review surface.

• Co-author of “A Multi-Agent Orchestration Framework for Explainable Human-in-the-Loop AutoML” — graph-orchestrated AutoML, SHAP/LIME explainability, human governance, and compliance-aware reporting
• YC Startup School India 2026
• IIT Guwahati Startup Expo 2026 — 2nd overall
• Anthropic Built with Opus 4.6 — finalist
• NASA Open Science — OS101 / OSE
• Mistral AI Global Hackathon

TypeScript Python JavaScript React Next.js TanStack Node.js Bun Hono FastAPI PostgreSQL Supabase Drizzle Docker PyTorch LLM APIs RAG Embeddings Agent Systems

LinkedIn · X · ORCID · Medium · ayushmanmukherjee12@gmail.com