I build agentic systems end to end and make them hold up in production: the context they work from, the evaluation that gates each change, and the controls on what they can do. I wrote Nexus-MCP, an open-source code-context server for AI coding agents, and the AI Hub that showed an engineering org what its AI spend was buying.
Nexus-MCP gives coding agents vector, BM25 and code-graph context. Published on PyPI and the official MCP Registry, 1,301 tests.
See Nexus-MCPA golden dataset and scope checks test this site's chat agent. Run them before a release.
Try the agentGDPR and EU AI Act audit at Move By Vision. Scoped agent skill for governed data at Lendable.
See the evidence
Shreyas Jagannath
Senior Applied AI Engineer · Context Engineering, Evaluation & Governance for Production Agents
Ask my agent
It answers from my profile only, says how you can check each claim, and lists the gaps. It runs under the same evaluation and guardrails I build for production agents.
Pick a question to put it in the chat, or write your own.
Paste a job description. You get a fit verdict (strong, partial or weak), each requirement mapped to evidence from my profile, and the requirements I cannot support listed as gaps.
Work
A published package, a public repo and a live product. Each card says how you can check it.
Open-source MCP server that gives AI coding agents hybrid code search: vector, BM25 and a code graph. Measured: 95% fewer tokens than Grep and Read to reach the right file on Chromium base/, and 22% lower agent cost on an unfamiliar 460-file project. On PyPI and the official MCP Registry, 1,301 tests, fully local.
Built to unblock a client engagement: a large, undocumented Flutter codebase that AI agents could not navigate reliably. Combines vector search, BM25 and a structural code graph (containment, imports and call edges), so an agent gets the relevant code instead of whole files. Measured: 95% fewer tokens than Grep and Read to reach the right file on Chromium base/, and 22% lower agent cost on an unfamiliar 460-file project. Published on PyPI (nexus-mcp-ci, v2.2.0) and the official MCP Registry, and listed on the Glama MCP marketplace. Anyone can check the fit on their own code with a free, offline eval (uvx nexus-mcp-ci eval) that calls no LLM and sends nothing. 8 tools in one local server, tree-sitter parsing across 25+ languages, about 90 MB idle and 460 MB with the embedding model loaded, no cloud dependency and no API keys.
AI coding agents spend most of their token budget reading irrelevant code — grepping for a function, reading entire files, repeating for every dependency. Existing code intelligence tools are fragmented across separate servers, each consuming 500 MB+ RAM.
Built a single MCP server consolidating two predecessor projects. Dual parsing (tree-sitter for symbol extraction, ast-grep for structural graph) feeds vector and BM25 search fused with Reciprocal Rank Fusion (vector 0.5 + BM25 0.3), with a code-graph boost for hub symbols. Added an ONNX Runtime model path (~50 MB vs ~500 MB on PyTorch) alongside the default bge-small-en model, used LanceDB with mmap for disk-backed vectors, and added token budgeting, semantic memory, call-edge resolution, and production hardening (graceful shutdown, corrupt index recovery, rate limiting, audit logging). v2.1.0 cut the tool list from 10 to 8 and removed FlashRank re-ranking, after benchmarks showed that fewer, richer tools and smaller results cost less. v2.2.0 added nexus-mcp-ci eval: a free, offline check that compares Nexus-MCP with Grep and Read on questions made from the user's own code, plus an opt-in paid tier that runs Claude Code or Codex with a spend cap. Every published benchmark number now comes from one results file, and CI fails when a generated block is out of date.
Measured: 95% fewer tokens than Grep and Read to reach the right file on Chromium base/, and 22% lower agent cost on an unfamiliar 460-file project. 8 tools, about 90 MB idle and 460 MB with the embedding model loaded — indexing, hybrid search, symbol lookup with callers and callees, a call graph with transitive impact analysis, persistent memory and architecture mapping across 25+ languages. 1,301 tests, 21 Architecture Decision Records, published on PyPI (pip install nexus-mcp-ci) and the official MCP Registry, listed on the Glama MCP marketplace, CI/CD with trusted publishing. Zero API keys, fully local.
How this was measured: Free offline eval (nexus-mcp-ci eval, no LLM, 2026-11-03): on Chromium base/ (3,638 files, 40 questions) Nexus-MCP reached the right file with a median of 950 tokens against 17,998 for Grep and Read, and put it first 70% of the time against 40%. On pallets/click (91 files, 30 questions): 1,301 tokens against 24,774, right file first 97% against 27%. Live agent runs (Claude Code with Claude Sonnet, 12 questions, 3 runs each, v2.1.0): on an unfamiliar 460-file Python project the cost fell 22% ($0.571 to $0.446) and turns fell from 55 to 35, with 35 of 36 answers correct. On Django, which the model already knows, the cost was about equal (−3%), turns fell from 93 to 71, and 36 of 36 answers were correct against 34. The same code varies by about 20% in cost over 3 runs.
Clinical AI platform that turns supplement queries into evidence-grounded recommendations. A Pharmacovigilance Critic checks every recommendation against the user's allergies and medications.
Clinical AI platform that turns supplement queries into evidence-grounded, safety-checked recommendations. A 9-node LangGraph pipeline grounds product data against NIH DSLD and PubMed in a Neo4j knowledge graph, with a Pharmacovigilance Critic checking every recommendation against the user's allergies and medications.
Supplement and wellness recommendations lack grounding in clinical evidence and fail to account for individual contraindications, creating real safety risks for users.
9-node LangGraph pipeline grounding product data against NIH DSLD and PubMed in a Neo4j knowledge graph. A Pharmacovigilance Critic node checks every recommendation against the user's declared allergies and medications. DeepEval quality gates enforce output standards; published healf CLI exposes the platform.
Evidence-grounded wellness guidance with built-in safety guardrails, contraindication checking, and DeepEval-validated output quality. Published as a CLI tool.
Describe a business intent; get a working, deployed prototype back in under 5 minutes. Six agent stages, with QA and review agents that request fixes before anything reaches the screen.
Describe a business intent; get a working, deployed prototype back in under 5 minutes. A hosted multi-agent tool that turns a described business goal into a running prototype autonomously — a 6-stage pipeline extracts domain context, spec and UX plan, generates a self-contained React component, then QAs and reviews it in parallel, self-healing through up to five correction rounds before anything reaches the screen.
Most prototyping tools produce generic wireframes — stakeholders see "Lorem ipsum" and "User A" and immediately check out, so a described business idea never gets a real gut-check before it's built.
Next.js/TypeScript multi-agent orchestrator and message bus: businessContext and spec agents extract domain entities, KPIs and BDD success criteria, a UX architect picks layout, charts and realistic domain-specific mock data, then a generator produces a sandboxed React component while QA and Reviewer agents run in parallel and request targeted fixes on failure. Built on DeepSeek V4.1 Flash, with a token budget and reasoning effort set for each agent role, to keep cost per run low. Hardened for a public demo: generated code compiles and runs in an opaque-origin sandbox iframe with its own CSP (no network, imports limited to React and Recharts), and a global LLM spend cap (runs a day and runs in flight) sits on top of per-IP limits, because per-IP limits alone do not stop an attacker with many addresses. Deployed on GCP Cloud Run via Cloud Build with runtime secrets from Doppler, with Vitest, Playwright and an evaluation suite.
A working, deployed prototype back from a plain-English business intent in under 5 minutes — wired with plausible domain data rather than placeholders. The evaluation suite scored 8.7/10 on average over 8 cases, and one run costs up to about $0.30 at peak prices.
How this was measured: Eval: npm run eval, 8 cases, 2026-10-07, low reasoning effort (high effort averaged 8.4/10 and was slower). Cost: DeepSeek V4.1 Flash peak prices, checked 2026-11-02.
3-agent LangGraph pipeline that turns a university curriculum into study guides, narrated video and quizzes. A two-layer evaluation framework gates every output.
A Python agent harness that connects to live data sources — databases, APIs, GCP infra — and builds validated dashboards, with every SQL path routed through one read-only guard by construction and no response model ever carrying a credential back to the LLM. In active development (MVP, Stage 3 of 5): the FastAPI gateway, SSE streaming and credential-hardening layer are built and tested (274 passing tests); the front-end and Cloud Run deploy are still ahead.
A daily pipeline that pulls roles from 13 ATS platforms and several job boards, checks each employer and salary against fixed rules, and scores what's left in two tiers: a free keyword pass first, then one LLM call only for jobs that survive it.
Built a Python-based NLP agent that converts raw consumer data into a Neo4j knowledge graph and structured vector store, feeding OpenAI, Gemini and Veo 3.1 video generation components with user-specific context for hyper-personalised video content.
Implemented an end-to-end agentic data pipeline to automatically crawl the web, qualify companies, enrich contacts and persist leads into an Enterprise GraphQL CRM.
How I build
Three things decide whether a production agent holds up. This is what I think about each, and where I have done it.
An agent is only as good as the context it gets. I model data, code and knowledge so that each agent sees the part it needs, and no more.
Nothing ships on trust. A golden dataset and a gate decide whether a change goes out, not a good demo.
Controls live in code, not in the prompt. Guardrails, read-only access and audit trails go in before release, not after an incident.
How I work: find the real problem with the people who have it, choose the trade-offs (guardrails, evaluation, data contracts), then ship with instrumentation and own the deployment path.
I build AI systems that stay inside limits enforced in code: guardrails on model output, evaluation gates before release, access controls on data, and audits against GDPR and the EU AI Act. The work below runs in healthtech, wellbeing and enterprise settings, where a failure has a real cost.
Agents act only inside limits that the code enforces. A prompt is not the control.
Nothing ships on trust. Automated evaluation scores each output, and a person keeps the final say where the stakes are high.
Secrets and personal data stay out of the model, the logs and the browser.
Compliance work is engineering work with an audit trail, done before release.
The right people see the right data, and every metric has an owner.
Experience
Each card gives the result first. The problem, the architecture and the full outcome are under Show details.
Lendable Operations Ltd
Jul 2026 — Sep 2026
Modelled engineering data into governed Snowflake metrics and gave engineers scoped, auditable agent access to it.
Klyft Technologies
Sep 2025 — Nov 2025
Took an eight-agent coaching platform to production on Google ADK and Vertex AI, with each agent scoped to its own data.
We Are Human Agency AB (Move By Vision)
Dec 2025 — Jun 2026
Rescued a failing healthtech AI platform in alpha and audited it against GDPR and the EU AI Act.
How this was measured: 100% of AI-induced crash failures removed, measured over the release in alpha-stage testing.
Scan and Sew Inc.
Oct 2024 — Aug 2025
Shipped a Designer AI Assistant (RAG and vision-language models over the design team's internal libraries, co-built with the designers and adopted into their daily workflow) and a real-time manufacturing tracking platform.
AIMAGE Technologies
Jan 2020 — Aug 2023
Cut p95 visual-search latency from 15s to 3–4s with knowledge graphs, and grew the team to 10.
Speaking
I talk about context for AI coding agents, and about evaluation and governance for production agents. London in person, or remote anywhere. Podcasts, meetups and conferences.
Engineers who use or build AI coding agents, and developer-tool teams.
Open-source maintainers, MCP builders and platform engineers.
AI platform, security and governance teams.
Before: Workshop at ODSC (Open Data Science Conference).
Background
Modern Machine Learning Approaches For Robotic Path Planning (Peer-Reviewed Journal)
Contact
I'm open to applied AI, context engineering and AI governance roles where I build and own a system inside one team. London, hybrid or on-site. For a podcast or an event, choose “Podcast or event invitation” in the form. The fastest route is email.