Senior Software Engineer · San Jose, CA

Ghanashyam Tatti

I build AI systems that have to be right — the tool layers, evaluation gates, and distributed services behind Cisco Catalyst Center's AI Assistant. My working theory is that a production LLM agent is a distributed software system, not a prompt: correctness lives in typed tool contracts, grounded context, deterministic response shaping, and evaluation that gates every release.

Eight-plus years across agentic AI, network analytics, and path computation at Cisco — and on nights and weekends I ship local-first software: an on-device AI notes app in both app stores, a calm task manager, and a shelf of open-source experiments.

8+
years shipping production systems
384
merged pull requests since 2021
812
merged changes reviewed
2
app stores shipping my on-device AI app
Selected work

Systems I own and ship at Cisco

The common thread: treating AI behavior as a software contract — typed tools, grounded context, and evaluation gates — and treating distributed state as something you design for, not hope about.

Reliable tool-using AI for network operations

Catalyst Center's AI assistant answers over live inventory, health, issue, and troubleshooting data — where a plausible-sounding answer can still use stale context, pick the wrong capability, omit a filter, or stop after one page of results. I moved that behavior into owner-correct boundaries: typed tool schemas, current-turn grounding, query builders, deterministic response shaping, and executable contracts, with fixes delivered across six workflow families.

  • LLM agents
  • MCP
  • typed tool contracts
  • Python

AI evaluation as a release control system

I own the evaluation and promotion gates for the shipped assistant: evaluations run daily, before my merges, and before twice-weekly promotions — and they catch regressions every week. Failures are traced through agent selection, tool calls, arguments, raw results, and delivery status rather than trusting an aggregate score. The offline replay platform I built is used by developers and testing teams and has supported resolving multiple customer-reported defects.

  • LLM evals
  • golden datasets
  • trace adjudication
  • release engineering

Scaling network analytics without weakening correctness

Network health APIs repeat expensive graph and metric queries under concurrent dashboard traffic. I delivered analytics APIs with query-level sorting and pagination, asynchronous execution, single-flight work sharing, and selective caching keyed by authorization-safe request identity — validated under concurrent load with every response equivalent to the uncached path.

  • Java
  • Spring
  • caching
  • concurrency

Owning Path Trace, a distributed path-computation service

Sole owner of the service that reconstructs how traffic moves through switched, spanning-tree, equal-cost, routed, and policy-controlled segments — on a platform whose largest documented profile supports up to 18,000 devices and 1.5 million ports. Architecture work spans site authorization, overlapping address spaces, IPv6-aware paths, on-demand ACL collection, and calculation-scoped spanning-tree state.

  • distributed systems
  • network protocols
  • Java
  • graph queries

Power telemetry, device to dashboard

An end-to-end energy analytics capability across four repositories: device telemetry parsing, streaming aggregation, graph-backed APIs, dashboards, and automated acceptance coverage. Allocation, consumption, and savings stay separate first-class metrics, with units, null behavior, time granularity, and late data carried through every layer.

  • event streaming
  • time-series
  • React
  • OpenAPI

A resilient device-security state lifecycle

Designed and shipped a dual-source lifecycle that detects when managed devices enter or leave a less-secure operating state — choosing redundant telemetry over a simpler single source to avoid missed transitions, then documenting and instrumenting the races that choice buys. The resulting security-state field is part of the public Catalyst Center API surface.

  • distributed state
  • idempotency
  • API design

Keeping issues and automated diagnosis trustworthy

Active operational issues could silently disappear at retention expiry while the underlying condition persisted. I built the scheduled lifecycle that verifies the resource still exists and republishes unresolved issues, and fixed automated root-cause workflows that reached false conclusions — a wrong source of truth here, a reused execution scope there. Encoding expert troubleshooting as repeatable inference is the product.

  • automated RCA
  • workflow engines
  • observability

Reproducible incident forensics

Led AI incidents that crossed orchestration, model, tool, and transport boundaries: provider throttling hidden by graceful fallbacks, successful root traces hiding failed child calls, completed graphs that never delivered a terminal event. I converted the recurring work into reusable methods and a secure trace exporter — paginate to exhaustion, verify integrity, protect credentials — now used in daily customer-environment triage.

  • distributed tracing
  • incident response
  • LLM observability
Projects

Things I build on my own time

ThinkAloud

Speak your mind. Voice notes → smart Markdown, 100% on-device AI.

A cross-platform notes app (iOS, Android, macOS, Windows) that turns voice, photos, and quick thoughts into structured Markdown — with quantized LLMs running entirely on the device via LiteRT-LM. No cloud, no accounts, no telemetry; semantic search over notes comes from a locally built embedding index.

  • Flutter
  • LiteRT-LM
  • Gemma
  • on-device AI

Cairn

A calm, private, local-first task manager.

Three projects, one next action each, a board of hard deadlines, and a fifteen-minute weekly reset. Offline-first PWA with cross-device sync; reads come from the local cache, writes go to D1, and nothing in the UI is ever red.

  • SvelteKit
  • Svelte 5
  • Cloudflare D1
  • PWA

Nala

A personal cookbook that treats recipes as code.

Create, version, and experiment with recipes — synced to GitHub with full history. WebLLM runs a model in the browser for nutrition calculation and plain-text recipe parsing, so personal data never leaves the device. Named for the ancient king who wrote the Pakadarpana.

  • React
  • WebLLM
  • GitHub API

Aegis

A CLI coding agent with a self-correction loop.

Plan → Code → Lint → Critic: every task goes through planning, implementation, tree-sitter-backed linting, and critical review before you see the result. Works with any LLM backend, ships a full REPL with file references, sessions, and safety guardrails.

  • Python
  • agents
  • tree-sitter

ResumeCrafter

An AI resume builder that tailors a master resume per application.

Maintain one comprehensive master resume, then let a client-side LLM tailor it to each job description — ATS-oriented, keyword-aware, and entirely in the browser.

  • JavaScript
  • client-side LLM
Babble A fault-tolerant, pub-sub social network. GitHub ↗
PokeGAN Generating new Pokémon with DCGANs. GitHub ↗
ReverbSynth Audio synthesis with effect interpolation. GitHub ↗
Experience

Where I've worked

2021 — present

Software Engineer — Catalyst Center

Cisco · San Jose, CA

  • Architected core Model Context Protocol tools and agentic pipelines for the AI Assistant platform, expanding system capabilities to drive an 80% increase in user engagement.
  • Designed core architecture for the Deep Reasoning agent's network troubleshooting and root-cause analysis, authored runbooks, skills, and tools for its agentic harness, and drove its beta program with 60+ enterprise customers end to end.
  • Led the AI Assistant's migration from GPT-4.1 to GPT-5.4-mini single-handedly: designed and ran the evaluation suite, then rewrote prompts, tool definitions, and agent instructions to hold answer quality while cutting inference cost.
  • Optimized deep-reasoning agent infrastructure by resolving critical bottlenecks, reducing overall latency by 40% and token consumption by 60%.
  • Owned the Path Trace diagnostic service end to end, built PoE power-telemetry analytics used across 70+ network deployments, and engineered an API-level caching layer that cut latency for concurrent requests by 70%.
2020

Software Engineer Intern — Catalyst Center Assurance

Cisco · San Jose, CA

  • Pioneered an anomaly-detection system for automated root-cause analysis of customer bugs, reducing process duration by over 80%.
2016 — 2019

Machine Learning Engineer — Data Science

ThoughtClan Technologies · Bangalore, India

  • Built a low-cost gas-leak detection system by training a CNN on infrared images — 85% detection accuracy at 98% lower cost than the hardware alternative.
  • Engineered an AutoML pipeline for industrial IoT: >90% accuracy with 10-second remote deployment of TensorFlow Lite models to mobile and edge devices.
  • Shipped a chatbot builder with over 96% intent/entity recognition accuracy, deployable as a standalone service or an Alexa skill.
2015 — 2016

Software Engineer

Rare Mile Technologies · Bangalore, India

  • Engineered social-media integrations, platform-wide search with autocomplete, and an XLSX import tool that onboarded hundreds of thousands of customers in seconds.
Skills

Tools I reach for

AI & agents
  • LLM agent architecture
  • MCP
  • tool-contract engineering
  • LLM evaluation
  • LangGraph
  • LangSmith
  • RAG & vector search
  • vLLM
  • PyTorch
  • TensorFlow
Languages
  • Python
  • Java
  • TypeScript
  • Dart
  • Go
  • SQL
Backend & distributed
  • Spring
  • FastAPI
  • Kafka
  • RabbitMQ
  • graph & time-series queries
  • Postgres
  • Elasticsearch
  • caching & concurrency
Apps & frontend
  • React
  • SvelteKit
  • Flutter
  • Cypress
  • Playwright
Cloud & operations
  • Kubernetes
  • Docker
  • AWS
  • GCP
  • Azure
  • Cloudflare Workers & Pages
  • Grafana
  • CI/CD & release promotion
Education

Where I studied

New York University M.S. Computer Science 2019 — 2021
Visvesvaraya Technological University B.E. Computer Science and Technology 2011 — 2015
Contact

Say hello

The fastest way to reach me is email. I'm always happy to talk about agentic systems, evaluation, local-first software, or an interesting problem you're stuck on.