Akram - RAGMohammed Akram Khan Lodi — CS undergraduate researching recursive language models, self-improving agent harnesses, and the energy cost of efficient AI.Zola2026-04-15T00:00:00+00:00http://akramlodi.com/tags/rag/atom.xmlNDA Analyzer: a Multi-Agent RAG System2026-04-15T00:00:00+00:002026-04-15T00:00:00+00:00
Mohammed Akram Khan Lodi
http://akramlodi.com/projects/nda-analyzer/<ul class="link-row">
<li><a href="https://github.com/akramlodi/NDA-Analyzer---multi-agent-orchestration">Code on GitHub</a></li>
</ul>
<h2 id="what-it-does">What it does</h2>
<p>Employment NDAs are long, templated, and easy to sign without reading. NDA Analyzer reads one for you and returns a clause-by-clause review: what each clause says, where it’s loose or one-sided, whether it holds up against Indian law and common HR policy, and how to rewrite it.</p>
<h2 id="architecture">Architecture</h2>
<p>The work is split across <strong>four agents that run in sequence</strong>, each with a narrow job and its own prompt:</p>
<ol>
<li><strong>Clause extraction.</strong> Segments the document into individual clauses and labels them.</li>
<li><strong>Loophole detection.</strong> Flags ambiguous, overly broad, or one-sided language in each clause.</li>
<li><strong>Legal-compliance verification.</strong> Checks flagged clauses against retrieved statutes and policies.</li>
<li><strong>Fix generation.</strong> Proposes concrete rewrites for clauses that fail.</li>
</ol>
<p>Splitting the task this way keeps each agent’s context small and its failures easy to localize. That same idea is what drew me to <a href="http://akramlodi.com/research/self-harnessing-rlms/">harness engineering for recursive language models</a>.</p>
<h2 id="grounding-with-retrieval">Grounding with retrieval</h2>
<p>The compliance and fix agents are grounded with a <strong>retrieval-augmented generation</strong> pipeline: Indian legal statutes, NDA templates, and HR policy documents are embedded with sentence-transformers and stored in <strong>ChromaDB</strong>, and the relevant passages are retrieved for each clause before the model judges it.</p>
<h2 id="stack">Stack</h2>
<p>Next.js and TypeScript on the front end, <strong>Gemini 2.5 Flash Lite</strong> for the agents, REST APIs between stages, and <strong>Server-Sent Events</strong> to stream each agent’s output as it’s produced.</p>