← All speakers

Sohail Shaikh

Conference affiliation: Prosodica · 2026

Sohail Shaikh is a data scientist working across analytics, natural language processing, conversational intelligence, and retrieval-augmented generation. His work on agent tool selection addresses a concrete scaling problem: as a tool catalog grows, sending every schema with every request can make the model slower, more expensive, and more likely to choose the wrong function. With Ankush Rastogi, he has explained how retrieving a small set of relevant tools can reduce that burden.

On this page

Sohail’s background spans engineering, marketing analytics, and applied AI. His public projects include A/B testing, customer-lifetime-value analysis in Python, and a car-rental database built with SQL: work concerned with comparing outcomes, estimating customer value, and organizing business data. At his 2026 AI Engineer appearance, he was a data scientist at Prosodica, focusing on applied AI, NLP, conversational intelligence, and RAG systems.

Retrieving tools instead of loading the catalog

In their joint presentation, Sohail and Ankush approached agent design from complementary directions: Sohail focused on model and routing behavior, while Ankush concentrated on system design, implementation, and production trade-offs. Their starting point was the convenience of giving a model every available tool definition. That approach can work for a small catalog, but unrelated schemas consume context and similar tools become harder to distinguish as the catalog expands.

Semantic routing applies the retrieval pattern used in RAG to tool selection. Tool descriptions are embedded and stored in a vector index when the catalog is created or updated. At runtime, the user’s request is embedded with the same model and matched against that index. Only the selected tools’ schemas enter the model call. Teams with existing document-retrieval infrastructure can reuse much of it for this purpose.

The pair combine retrieval with just-in-time context injection: wait until the request is known before loading the capabilities it needs. This reduces the model’s choice set as well as its input tokens. The benefit depends on retrieving the right tools; a necessary function excluded by the router is unavailable to the model for that call.

Evaluating the smaller choice set

Their approach makes several decisions explicit:

  • Write descriptions for retrieval. Descriptions should use the language users employ and identify the action and relevant entities. A tool can remain difficult to retrieve even when its schema is valid, so repeated misses are a reason to revise its description.
  • Measure the candidate-set trade-off. They compare tool-selection accuracy, time to first token, input tokens, and estimated cost while varying the number of retrieved tools, K. Smaller sets reduce context; larger sets can recover missed candidates. They propose five as a starting point, then testing three, five, and ten against the application’s accuracy target.
  • Provide a recovery path. If the selected tools cannot support the task, widen the candidate set, run another retrieval pass, or search a broader tool group. Logging retrieved candidates and final tool calls helps distinguish a retrieval miss from a model-selection error.

Sohail and Ankush also preserve a place for direct loading. A system with ten or fifteen tools may gain little from a router. Their recommendation is to introduce retrieval when catalog size produces measurable problems with context, latency, or tool confusion, and to tune it against those problems.

Read the topics behind these talks

1 conference talk

Loading more talks and highlights…

References