Sohail Shaikh is a data scientist specializing in applied AI, natural language processing, conversational intelligence, and retrieval-augmented generation. His work addresses a practical problem in agent design: loading every available tool into every request makes models slower, costlier, and more likely to choose incorrectly.
His technical background spans engineering, analytics, and experimentation. Public projects include A/B testing, a SQL car-rental database, customer-lifetime-value analysis, and odsc-lancedb.
At AI Engineer World’s Fair 2026, Shaikh represented Prosodica alongside Ankush Rastogi in a session on scalable agent tooling. Their central ideas include:
- Semantic routing as RAG for tools: Embed and index tool descriptions, match each request against that index, and provide the model only the most relevant schemas.
- Just-in-time context injection: Load capabilities after understanding the request, keeping the active context compact even as the overall tool catalog expands.
- Production-minded agent evaluation: Measure tool-selection accuracy, first-token latency, and token costs together; address retrieval misses through stronger descriptions, adjusted retrieval limits, and fallback searches.
Shaikh favors proportionate implementation: smaller toolsets can remain statically loaded, while larger catalogs benefit from retrieval infrastructure that limits irrelevant choices.
