Nilesh Sarkar / Internship

How the role evolved

The role started as a two-month IT Development - AI Engineering internship in Generative AI, focused on getting LangChain and LangGraph stacks running on Moog's air-gapped infrastructure. The scope kept expanding. Based on delivery, I was extended twice - both extensions full-time, Monday to Friday, on-site at MITC Bengaluru. Over the full run it stopped feeling like an internship and started feeling like research-and-engineering work: not just building agents, but reading the underlying papers, understanding the failure modes, and figuring out what it actually takes to make AI reliable in production.

Each project in the timeline below corresponds to a real, deployed system used by aerospace engineers at MITC - measured numbers, not demos.

Deployment context & constraints

Moog Controls builds safety-critical hardware for aerospace and defence, so every system I shipped had to operate inside three hard constraints simultaneously:

• Air-gapped, on-network. Nothing could leave the corporate intranet. No external API calls, no managed LLM endpoints, no cloud RAG vendors. Models, vector stores, agents, and traces all run on-premises.
• Regulated, audit-friendly. Aerospace and defence-grade engineering documentation needs reproducible provenance. Retrieval traces, agent step logs, and answer sources are recoverable for every query.
• High reliability bar. The target was 99%+ uptime on systems used by 100+ engineers in production workflows. Failures had real engineering-throughput cost, not just inconvenience.

These constraints shaped every architectural choice below - choice of LangGraph over a single-shot pipeline, the decision to build an internal MCP server instead of relying on tool-calling alone, and the move to a CV-only auth path for the restricted lab.

Project timeline

• Phase 1 - RAG foundation. Air-gapped LangChain RAG over aerospace engineering corpora; baseline retrieval at 70%.
• Phase 2 - Agentic RAG. Moved to LangGraph; multi-step retrieval, re-ranking, query rewriting. Retrieval climbed to 90%+.
• Phase 3 - Multimodal document intelligence. Agents that parse engineering diagrams, embedded images, and dense technical PDFs end-to-end.
• Phase 4 - Internal MCP server. Standardised tool exposure for LLM agents across engineering workflows; 37% productivity gain measured.
• Phase 5 - Computer-vision gesture authentication. Hands-free login + equipment control for restricted-lab environments. 2-month deployment, 99.99% uptime.
• Phase 6 - Supply-chain agents. Procurement, document routing, status tracking across 3 departments. 2.5x workflow efficiency.
• Phase 7 - LLM-assisted CAD prototype. Natural language to parametric CAD geometry with real-time FEA; spun out as an independent research project.

Key Contributions

• Agentic RAG Pipelines (+ custom UI)
  Deployed air-gapped LangGraph + LangChain + LangExtract agentic RAG pipelines for aerospace engineering workflows; iterated on retrieval accuracy, response confidence, and latency until the system was trustworthy enough for engineering use on complex technical documents. Retrieval accuracy went from 70% to 90%+; built a custom UI around the system so adoption wasn't a barrier. Electronics-department agent suite achieved 55% productivity boost, 94% adoption (100+ users) across 3-4 deployed agents, 99.3%+ average reliability.
• Multimodal Document Intelligence
  Extended the RAG system to handle multi-modal documents: flowcharts, engineering diagrams, embedded images, and dense technical PDFs all in one unified knowledge base. Lets engineers query mixed-media sources from a single place rather than chasing the right document type each time. All data stays on-network.
• Internal MCP Server
  Built and deployed an internal MCP (Model Context Protocol) server exposing company tools as structured interfaces for LLM agents, automating engineering workflows; 37% productivity gain. Deployed the Verilog MCP Server for Verilog engineering workflow automation.
• Gesture Recognition System
  Built a computer vision gesture recognition system for restricted lab environments: hands-free authentication and equipment control - 98.6% accuracy, 6x faster login, 1.4% FRR, 99.99% uptime (2-month deployment), 97% adoption.
• LLM-orchestrated CAD & Automated Engineering Analysis
  Designed a system where an LLM acts as orchestrator over a deterministic CAD engine to generate manufacturable 3D geometry from natural language. On top of that, an automated engineering-analysis layer runs stress, strain, and volumetric simulations on existing CAD models, auto-generates design variations, and tests them against specified tolerances to surface the optimal model. Proof-of-concept complete, presented to the CAD design team. Continuing as an independent research project on parametric text-to-CAD.
• AR / VR Hand-Tracking Exploration
  Separate from the lab-auth gesture system: a computer-vision hand-tracking exploration for AR / VR use cases, inspired by mixed-reality interaction patterns from devices like the Apple Vision Pro. Same CV foundations as the gesture-auth work, different downstream use case.
• AI Agents for Supply Chain
  Built and deployed AI agents for supply chain operations across 3 departments - autonomously handling procurement queries, document routing, and status tracking; 2.5x workflow efficiency improvement.

Technologies & Impact

Retrieval accuracy improved from 70% to 90%+. Electronics department agent suite achieved 55% productivity boost with 94% adoption (100+ users) across 3-4 deployed agents. Computer vision gesture system reached 98.6% accuracy with 99.99% uptime (2-month deployment). Supply chain AI agents delivered 2.5x workflow efficiency improvement across 3 departments. MCP server integration measured 37% productivity gain for engineering workflows.

Engineering choices & lessons

• LangGraph over single-shot RAG. A flat retrieve-then-answer pipeline plateaus around 70% on aerospace docs because the query language is dense and the right chunks rarely sit in the top-k by cosine alone. LangGraph let me route through query-rewriting and re-ranking steps with explicit state, which is what pushed retrieval past 90%.
• MCP server, not raw tool-calling. Once three engineering teams wanted agent access, ad-hoc tool wrappers stopped scaling. Building an internal MCP server gave every team a single, versioned, schema-typed entry point - which is also what made the 37% productivity number measurable rather than anecdotal.
• CV-only auth for restricted labs. The lab couldn't accept any input device that touched a network. A camera with on-device inference (no remote calls, no biometrics leaving the device) was the only architecture that cleared the security review; the 1.4% FRR target came directly out of that constraint.
• Reliability before features. The 99.3%+ and 99.99% uptime numbers were not free - they came from explicit retry-and-fallback logic, on-network embedding cache invalidation, and a fail-safe-to-last-known-good rule for every agent step. Adding a feature meant adding the failure mode for it first.