Speaking: Nexus-MCP, AI agents and governance

Talk topics

Talk 1

Your coding agent reads too much

For: Engineers who use or build AI coding agents, and developer-tool teams.

AI coding agents grep, open whole files and read the same dependency again and again. Nexus-MCP started when agents could not find their way around a large, undocumented Flutter codebase on a client project. This talk shows how vector search, BM25 and a code graph, fused with Reciprocal Rank Fusion, let an agent find relevant code with an estimated 70–90% fewer tokens, and where that estimate stops.

The audience leaves with

Why one retrieval method is not enough for code.
How to weight and fuse vector, BM25 and graph results.
How to state a token saving honestly.

Talk 2

Shipping a local-first MCP server that people install

For: Open-source maintainers, MCP builders and platform engineers.

Nexus-MCP runs on the developer's machine with no API keys and no cloud dependency. It fits 10 tools in a 350 MB memory budget, has 607 tests and 18 Architecture Decision Records, and ships to PyPI with trusted publishing. The talk covers the trade-offs: an ONNX Runtime model path at ~50 MB instead of ~500 MB on PyTorch, disk-backed vectors in LanceDB, and what the download data does and does not prove.

The audience leaves with

How to keep an MCP server inside a memory budget.
Which decisions earn an ADR in a small open-source project.
What a PyPI download count includes, mirrors among them.

Talk 3

Governed context for agents in regulated settings

For: AI platform, security and governance teams.

At Lendable, an FCA-regulated lender, engineers needed agent access to governed engineering data. A scoped, auditable agent skill met the security requirements, so an MCP server was not used. In Agent Harness, every SQL path goes through one read-only guard, and adversarial review found a ReDoS in the redaction regex (2.4s down to 8ms). This talk is about controls that live in code, not in the prompt.

The audience leaves with

When a scoped skill is safer than an MCP server.
Read-only by construction, not by convention.
What adversarial review finds that unit tests miss.

Talk 4

Evaluating the agent on my own website

For: Teams that put a first LLM agent in front of users.

This site has a chat agent and a JD Match analyzer. The chat agent works under a scope lock that a seven-case scope eval tests, and a ten-case golden dataset checks its answers. JD Match must trace each claim to a profile entry and list the requirements it cannot support as gaps. A small, public example of evaluation gates and guardrails on a live agent.

The audience leaves with

How to write a scope lock and test it.
Evidence tracing as a guard against invented claims.
What to put in a golden dataset for a small agent.

Formats: Podcast · Meetup talk · Conference talk · Workshop · Panel.

Questions a host can ask me

Why build Nexus-MCP instead of using an existing code-search tool?
What does context engineering mean in practice for a coding agent?
How do you know that a token saving is real?
When should an agent get an MCP server, and when a narrower skill?
What does an evaluation gate look like for a small team?
What did founding a company teach you about shipping AI?

Bios to copy

Short bio · 52 words

Shreyas Jagannath is an applied AI engineer in London. He builds production agents and the context, evaluation and governance around them. He wrote Nexus-MCP, an open-source MCP server that gives AI coding agents hybrid code search. He has 8+ years of building software, including 3.5 years as founder and CTO of AIMAGE.

Medium bio · 110 words

Shreyas Jagannath is an applied AI engineer in London. He builds production agents and the context, evaluation and governance around them. He wrote Nexus-MCP, an open-source MCP server that gives AI coding agents hybrid code search. He has 8+ years of building software, including 3.5 years as founder and CTO of AIMAGE. At Klyft he took an eight-agent coaching platform to production and cut tokens on the heaviest agent messages 10x. At Lendable, an FCA-regulated lender, he gave engineers governed data access through a scoped, auditable agent skill. He holds an MSc in Artificial Intelligence (Distinction) from the University of Surrey and ran a workshop at ODSC for 100+ professionals.

Long bio · 199 words

Shreyas Jagannath is an applied AI engineer in London. He builds production agents and the context, evaluation and governance around them. He wrote Nexus-MCP, an open-source MCP server that gives AI coding agents hybrid code search. He has 8+ years of building software, including 3.5 years as founder and CTO of AIMAGE. At Klyft he took an eight-agent coaching platform to production and cut tokens on the heaviest agent messages 10x. At Lendable, an FCA-regulated lender, he gave engineers governed data access through a scoped, auditable agent skill. He holds an MSc in Artificial Intelligence (Distinction) from the University of Surrey and ran a workshop at ODSC for 100+ professionals. Nexus-MCP started on a client project: a large, undocumented Flutter codebase that AI agents could not navigate. It combines vector search, BM25 and a code graph, runs fully local with no API keys, and is published on PyPI with 607 tests. Shreyas started in computer vision research for ISRO's Chandrayaan rover, then built full-stack and cloud systems as a founding engineer and consultant. He has audited an AI healthtech platform against GDPR and the EU AI Act, and he writes guardrails and evaluation gates into the agents he ships.

Past speaking

Workshop at ODSC (Open Data Science Conference). Implementing AI and deep learning in enterprise environments, for 100+ professionals.
Best Presentation Award. Listed with the other awards on the resume page.

Invite me

Tell me the show or the event, the date, the audience and the topic you have in mind. I read every message and reply by email.