Ayush Manish Agrawal. Curious first. Then I build it.
I work on production multi-agent systems for cloud infrastructure operations at Microsoft. The best way to see what I like to build is to meet one: my agent is right below. Ask it anything about me and watch it dig up the receipts. Sound on is more fun.
Screen Ayush.
About 6 minutes · no score · you keep the notes
Agenda · Background · Experience · Projects · Fit check · Anything else · follow it or skipYour pick is remembered. Switch mid-session with one control.
Leads FIDGET, the shared knowledge graph behind CIS incident automation, and is building Vigil, a platform for always-on agent loops. Architected and built ARLI (CIS Copilot), a multi-agent Text-to-SQL system, and its evaluation.
Details
FIDGET: designed, architected and leads the knowledge layer behind CIS incident automation, a graph that holds 2.4M nodes and 6.2M links covering 420K+ incidents over a rolling 90 days, in production on AKS and served to every incident agent through one read-only retrieval API and MCP tools.
FIDGET's design: graph traversal and vector similarity in one query path, scoped to the owning team first; team knowledge federated in one graph (any team onboards by editing config, not code); ingestion as idempotent, single-writer scheduled jobs under a versioned schema.
Evaluated LLM root-failure extraction on 150 real task logs: an LLM judge preferred it over regex 77% to 4%, and 97% of its quotes matched the log verbatim. Agent findings stay suggestions until quote-verified and engineer-confirmed.
Vigil: building a platform on AKS for agent loops that run 24/7. Its first loop watches 15 services (about 17,800 telemetry checks a day) and has triaged 341 problems into 103 real issues and 223 noise, producing 42 fixes and PRs for $359 in total spend.
Built a labelled set of 529 anomalies across 37 root causes and compared six policies for which fixes ship immediately: a weighted scoring formula cut false-critical fixes from 18 to 1 while missing 1 of 4 critical cases, where LLM judges missed 2 or 3.
ARLI (CIS Copilot): architected and built it end to end (orchestrator, query, writer and visualization agents) so planners, SREs and leadership can query buildout status and incident impact in natural language; about 80% positive user feedback in production.
Introduced table rules, per-table constraint files written by domain experts, raising LLM-judged execution accuracy from 3.4% without tools to 54.5% on 300 production questions; measured with a 2⁴ full-factorial ablation and Shapley attribution. First author of the resulting paper (MSJAR 2026).
Built ETA prediction for GPU buildouts with HDBSCAN clustering and XGBoost regression: 79% of predictions within ±7 days, better than SLA-based estimates in 83% of cases, refreshed at sub-hour latency as incidents land.
Collaborates across teams on Microsoft Scout, Microsoft's always-on enterprise AI agent, extending its autonomous capabilities.
2019 — 2022
Software Engineer · University of Nebraska
Software Engineer (previously Programmer/Analyst). A full-stack .NET and React issue tracker, and a Python library that automated versioning and tagging on merge requests.
Details
Built a full-stack issue tracker (.NET Core, React) that raised user issue reporting by 55%.
Built a Python library that automated versioning and tagging on merge requests.
2020 — 2022
Research Engineer · Manifold Computing · OpenMined
Open research on how deep networks learn and on federated learning. Contributed talks at NeurIPS 2020 workshops.
Details
Manifold Computing (Sep 2020 – Sep 2022): designed a relative weight change (RWC) metric to interpret layer-wise learning dynamics in deep networks and ran the PyTorch experiments across standard computer vision architectures. First author of the resulting paper: contributed talk at the NewInML workshop at NeurIPS 2020, later published by Springer (2021).
OpenMined (Jul 2020 – Dec 2020): studied ML systems from a data-privacy perspective, reproduced federated learning algorithms in PyTorch, and proposed a metric for the complexity of federated learning systems (contributed talk).
Research
03 — Papers
MSJAR 2026 · Microsoft MLADS 2026 talk and poster
Grounding Enterprise Text-to-SQL in Tribal Knowledge
Introduces table rules, retrievable per-table constraint files written by domain experts, as a complement to schema and few-shot retrieval in Text-to-SQL agents. They raised LLM-judged execution accuracy from 3.4% to 54.5% on 300 production questions.
Hiring for applied AI or agent systems? Reach out with the role and I’ll reply myself. I’m also open to collaborations on agent systems, side projects that need a builder, research, and the occasional talk. I read everything.
