AC
BUILDING MACHINES THAT UNDERSTAND | TURNING DATA INTO DECISIONS | TEACHING SYSTEMS TO FIND THE SIGNAL | AI × DATA × CREATIVITY | FROM PROBLEM → SYSTEM → SOLUTION | CODE IS THE TOOL. THINKING IS THE ENGINE. | ENGINEERING INTELLIGENCE, NOT JUST SOFTWARE | WHERE LOGIC MEETS STORYTELLING | BUILD. BREAK. ITERATE. REPEAT. | CURIOUS BY DEFAULT. | ALWAYS BUILDING SOMETHING.

Case Study

SecureDocAI

An offline-first document intelligence system for retrieval, summarization, and evidence-aware question answering.

LLM/RAGIn Progress2026

Overview

SecureDocAI explores how privacy-sensitive documents can be queried locally using retrieval workflows and compact language models without cloud dependency.

Problem

Knowledge workers handling sensitive documents often need answer retrieval, but cloud-hosted systems can violate policy or trust boundaries.

Why It Matters

A private-by-default assistant unlocks faster decision making while preserving compliance and data ownership in restricted environments.

Role

Designed retrieval architecture, implemented document ingestion flow, and iterated on answer quality strategy.

Architecture

Pipeline: document parsing, OCR fallback, semantic chunking, embedding indexing in FAISS, retrieval layer, and offline response generation.

Architecture Signal

  1. Document

  2. Parsing

  3. OCR

  4. Chunking

  5. Embeddings

  6. FAISS

  7. Retrieval

  8. Offline LLM

  9. Answer

SecureDocAI Architecture

End-to-end offline document intelligence pipeline.

Documents
Parsing
OCR
Chunking
Embeddings
FAISS
Retrieval
Offline LLM
Answer
Evaluation

Stack

Python, FAISS, sentence-transformers, FastAPI, Docker

Implementation

Built ingestion modules with traceable metadata and a retrieval layer tuned for relevance and citation support. Added modular service boundaries for testing each stage independently.

Engineering Decisions

  • Prioritized local model compatibility before advanced generation features.
  • Used vector index snapshots to make iteration reproducible across experiments.
  • Separated retrieval scoring from response rendering to improve debuggability.

Challenges

  • Uneven OCR quality across document formats.
  • Retrieval drift when chunk sizes were too broad.
  • Balancing latency with answer depth on consumer hardware.

Solutions

  • Added OCR fallback and preprocessing normalization.
  • Introduced chunk overlap and metadata-aware retrieval filters.
  • Applied staged retrieval with capped context windows.

Results

The system now returns grounded answers with source context, making internal document lookup significantly more actionable for early testing scenarios.

Evaluation

Evaluation is currently based on curated question sets and relevance review. Quantitative benchmarking is planned with a larger validation corpus.

Metrics

Answer grounding score

TODO

Pending formal retrieval evaluation baseline.

Median retrieval latency

TODO

Will be captured after final model and index lock.

Lessons

  • Retrieval quality is often a larger lever than model size.
  • Traceability fields are essential for debugging hallucination-like behavior.

Future Improvements

  • Add reranking stage for improved top-k relevance.
  • Introduce multilingual document support.
  • Automate regression checks for retrieval quality changes.