Back to projects
AI / ML

AI Agent for Semiconductor Manufacturing

  • Sep 2025 - Present
  • Research Lead
  • Research

Directed the research and development of an AI agent for a global semiconductor manufacturer. The agent leverages RAG architecture to combine real-time Manufacturing Execution System (MES) data with historical quality parameters, enabling engineers to perform natural-language root cause analysis for packaging defects.

QUERYMES + VECTOR DBANSWER[1][2][3]LANGCHAIN · RAG · MES ROOT-CAUSEAI Agent for Semiconductor Manufacturing

RAG

Over MES + quality data

Hours → minutes

Root-cause analysis cycle

Natural-language

Query interface

01

The Problem

Root cause analysis for backend packaging issues was a manual, time-intensive process requiring engineers to cross-reference multiple data sources and MES logs, often taking hours to identify the source of quality deviations.

Process engineers were stitching together MES logs, recipe tables, and tribal knowledge by hand. The agent had to be useful in the first message, or it would lose them on the second.
02

The Solution

Built a LangChain/RAG-based agent that ingests and indexes real-time MES processing data alongside quality parameters. Trained models to correlate process variables with defect patterns, enabling conversational root cause analysis through natural language queries.

03

Key Decisions

  1. Retrieval over fine-tuning

    MES data shifts continuously — recipes, equipment, calibration. Fine-tuning would have stalled in re-training cycles. A retrieval layer over a freshly-indexed snapshot keeps the agent current without retraining, and lets engineers see exactly which records grounded the answer.

    Tradeoff: Heavier on indexing infrastructure, but trust came from showing the sources, not from a glossier model.

  2. Structured tool calls for live MES queries

    Beyond retrieval, the agent exposes typed tools to query MES tables directly — yield by lot, parameter drift over a window, recipe diffs between lots. The model picks a tool and arguments; the runtime executes; the answer cites the rows.

    Tradeoff: More schema work to expose tools safely, but kept the agent grounded in real, current data rather than hallucinated correlations.

  3. Citations in every answer, refusal when retrieval is thin

    Every answer ships with citations to MES rows or quality docs. If retrieval similarity falls below a threshold, the agent refuses and asks for a narrower question rather than guessing. Trust was the gating factor for adoption.

    Tradeoff: Higher refusal rate on vague questions, but engineers preferred a confident "I don't know" over a confident wrong answer.

04

The Impact

Significantly reduced root cause analysis time for backend packaging issues, enabling faster corrective action and improving overall yield in semiconductor manufacturing.

With hindsight

Next iteration: a feedback loop where engineers mark answers as useful / wrong / missing-context, feeding both the retrieval re-ranker and the answer evaluator. Without that loop, quality plateaus at whatever the initial prompt is good enough for.

Built with

PythonLangChainPineconeGPT-4PostgreSQLAWSDocker