🌐 Portfolio → chandananandi.github.io — the full story, with live results, architecture diagrams, and honest limits.
I build production AI systems for healthcare at Solix Technologies: multi-agent orchestration, RAG pipelines, LLM fine-tuning, and the backend that holds it together. That work lives on company infrastructure, so it is not here.
What is here is the work I do to think through hard problems on my own time, mostly around safe AI for network operations. These projects share a thesis I care about: an LLM should propose, a formal tool should verify, a human should approve, and the system should prove that what it claimed would happen actually happened. AI where it earns its keep, determinism where determinism matters.
A few I would point you to first:
-
verifiednet-arena (live dashboard) — a safety benchmark for AI network-operations agents, and the clearest statement of the thesis above: the agent proposes a fix, a deterministic scorer checks seven safety properties, approval binds to a sha256 of the exact plan, and a gated harness executes that plan in a Docker lab so the lab's own validator proves recovery. 13 implemented benchmark tasks span EVPN/VXLAN, BGP route-policy, SONiC ACL/SAI, and a cross-domain ambiguity task. On the live-backed EVPN tasks, 18/18 combined agent remediations were live-verified end to end in Docker FRR labs (suite_v2 plus the task-006 route-target task); the first BGP task adds a separate 3/6 clean-room live-verification result. On hard-mode tasks where the right move is to refuse, both evaluated models abstained 24/24 with 0 unsafe-action attempts. Every scenario, submission, and transcript is frozen under 20 sha256 artifact locks, with a Docker-free CI re-verifying it all. Honest scope: local Docker (FRR) labs only — not physical hardware and not production networks — and the SONiC ACL/SAI tasks are fixture-replay only.
-
closcall — evidence-grounded incident command for a containerized SR Linux datacenter fabric, and the fullest expression of the thesis above. A 312-incident fault corpus collected under live traffic shows classical detection is structurally blind to gray faults (AUROC 0.500) while temporal learned models recover them (~0.91) — every published number generated from one content-hashed run. Remediation is human-approved through a browser console whose approve button and executor share a single digest-bound gate: the UI cannot claim more safety than the executor enforces, and a test proves there is no side door.
-
sonic-intent-agent — natural language to verified network config on a SONiC switch, with a local LLM, Batfish pre-apply verification, and structural post-apply checks. Built in seven phases, each gated by a working demo.
-
neuronoc-network-ops-assistant — a safety-first NetOps console: anomaly detection, a LangGraph analysis workflow, LLM root-cause explanation, and admin-approved remediation that is plan-only by construction. No execution path exists in the code, and a CI scan enforces it.
-
evpn-vxlan-frr-lab — a containerized leaf-spine fabric: eBGP underlay, BGP-EVPN type-3 control plane, VXLAN data plane, with a Python validator on FRR and Docker.
I came to AI from Aircraft Maintenance Engineering, a field where systems fail loudly and safety is not optional, then an MS in Data Science. That background is why my projects are honest about their limits and built to be verified rather than trusted.
Open to conversations about healthcare AI, agentic systems, and ML infrastructure.
- MS Data Science, University of the Pacific
- 📍 Dallas–Fort Worth, TX