Jayesh Ranjan Kesari

Agent runtime and tools

• ·Wrote the agent loop that runs every AI query: tool schemas, capped iterations, result truncation, and a last tool-free call so the model has to answer in text.
• ·Set up a tag-based tool contract where tools declare what they do and what they need, so adding one takes no runtime changes.
• ·Removed an extra billed round on capped turns by offering file tools only when the prompt actually asks for a file.
• ·Added cooperative cancellation across the pipelines after finding that stopped queries were still writing their results.

Multi-provider model layer

• ·Built a shared library and model registry: 17 models across 6 providers, picked per request instead of baked into the pipeline; 86 releases in three months.
• ·Got prompt caching working with ordered breakpoints, bringing repeat tool-loop rounds down to a fraction of the first round's cost.
• ·Tracked down a content-filter block that only showed up through the proxy, then confirmed the fix on live calls: blocked calls went from 3/12 to 0/12.

Token accounting

• ·Built a per-user cost ledger that meters every paid call (chat, embeddings, vision, OCR, images) into SQLite with a ClickHouse mirror.
• ·Made missing usage and unpriced models record as such rather than as zero, so nothing free gets confused with something never measured.
• ·Read token counts straight from provider payloads so cached-input and reasoning tokens are not lost.

Documents and spreadsheets

• ·Routed uploads on text density rather than file extension, so image-only decks go to vision transcription and text PDFs to the normal extractor.
• ·Swapped naive chunking of spreadsheets for SQL over detected tables with schema-card embeddings, then made table discovery progressive so it fits the tool-result cap. Graded accuracy went from 2/8 to 6/8.
• ·Sandboxed model-written SQL: no external access, read-only statements, scoped to the current conversation's tables.

Connectors, tools and security

• ·Architected the MCP tool server: 49 tools, including 13 per-user connectors for Microsoft 365, Power BI, Google Workspace, Zoho and Databricks (the last two not yet run against a live account), passing delegated tokens as transport headers so the model can never set one.
• ·Designed host-side approval for outbound mail: the parked call is the card, and Send runs it with a fresh token and no model in the path. A teammate built the first version; I added replies, attachments and per-provider routing.
• ·Wrote most of 60+ test modules covering wire contracts, cancellation, metering and citations.

Earlier at Netision

• ·Text-to-SQL analytics and root-cause agents over a read-only warehouse; cut a root-cause turn from 77.8 s to 12 s with the report unchanged.
• ·Backends for an on-prem observability platform in Go and FastAPI: agents, a collector, a probe, and a control plane on Redpanda and ClickHouse.

• ·Built LLM agents with LangChain, CrewAI and the OpenAI API, where I moved from web development to agents.

• ·Front-end and back-end work on web applications.