Associate SDE · Kyndryl · building Agentic AI on Cloud
|
I am an Associate SDE at Kyndryl, fresh out of VIT University (CGPA 8.95, AIR 56 in VITEEE), engineering agentic AI systems on Azure AI Foundry for enterprise clients.
I live at the intersection of machine learning, cloud architecture and developer experience — from LSTM grids predicting power loads across 500+ households, to a multi-tenant agentic observability platform (A360) ingesting millions of agent traces, to a LangGraph governance agent that turns plain-English policies into enforced cloud pipelines.
Certified across AWS, Microsoft Azure, GitHub and Founderz. Always certified, sometimes asleep.
A360 is Kyndryl’s platform for watching how enterprise AI agent fleets actually behave in production. Executive dashboards, KPI drill-downs, embedded Kibana, agent trust scores, multi-persona dashboards, LLM FinOps, and a conversational interface over telemetry — running on a streaming pipeline from OTEL → Postgres → Azure Data Explorer → Elasticsearch → Next.js.
Composite reliability + quality metric, fed by composite ES indices (run-summary, decision, run-heatmap). Surfaced at /agent-trust.
20 insight types — cost spikes, latency regressions, model drift, error clusters — auto-curated and rendered in the UI.
Boarding function (Flex Consumption) provisions per-tenant ADX DB, Cosmos containers, ES index templates and registers the tenant.
A360 emits its own traces into A360 under the kyndryl tenant. Eat your own observability.
Frontend → Azure AD token → API gateway validates against JWKS, checks authorized_users in Cosmos.
Hourly views with 70-min overlap, daily with 25-hour overlap. SHA-256 doc IDs from natural keys.
A multi-agent LangGraph platform that takes a plain-English policy like
“All GCS buckets must have cost_center, environment and team labels”
and turns it into compiled rules, live drift detection, risk-scored
gcloud remediation plans, human-in-the-loop approvals, and audit-ready
evidence — running against real GCP (Gemini 2.5 Pro on Vertex AI) or fully
offline in mock mode for CI.
NL → structured PolicyConstraint via few-shot LLM. Rejects greetings, off-topic, or ambiguous policies.
Queries live GCS / Compute / IAM / SQL state; LLM compares to policy and emits structured violations.
Generates ordered, reversible gcloud commands per violation, with rollback hints and blast-radius estimates.
Guardrails first (BLOCKED / HIGH_RISK), then LLM rubric scoring 0–10. Auto-approve below 4.
Runs approved commands via gcloud, re-queries the resource for post-state proof, refuses unapproved plans.
Conditional routers short-circuit on rejection or cloud-error; per-stage timings let the UI render a Gantt of agent execution.
BLOCKED patterns force score=10. HIGH_RISK patterns boost LLM score by +6 to push past auto-approve.
7+ YAML controls for GCS, Compute, IAM, Cloud SQL, VPC. Loadable via /api/templates.
Mock / LiteLLM / Vertex AI for LLM. Mock / GCP for cloud. AWS & Azure scaffolded next.
EvidencePack bundles: policy + finding + plan + before/after state + execution log. Persisted to Firestore.
LangGraph · LangChain · FastAPI · Pydantic v2 · React + Vite · APScheduler · Cloud Build · Firestore · Podman.
Designing & shipping Agentic AI POCs and enterprise use cases on Azure AI Foundry. Delivering mission-critical insights via Kyndryl Bridge with predictive modelling and observability. Owner of multiple A360 services and the P2P hackathon-to-production track.
Engineered cloud architectures that improved scalability +40%. Orchestrated migration of business ops to the cloud.
Led 3+ initiatives, ran 2 hackathons and 4 coding challenges, impacting 500+ students.
5 clusters · nearest skill follows your cursor · hover to inspect · switch view →
tilt the cards · star · external links open the repo.
Multi-tenant agent observability with 31 ADX MVs synced to ES, FastAPI gateway, Next.js + Shidoka frontend, synthetic telemetry generator.
5-agent LangGraph platform that compiles natural-language policies into enforced GCP guardrails with drift fix, risk scoring, and audit evidence.
Python toolkit that finds repeated overlay patterns across hostnames and quantifies business impact (cost / brand value).
End-to-end ELK pipeline that cleans, indexes, transforms and visualises incident-ticket data — fully script-driven and reproducible.
End-to-end IT asset analytics pipeline — pandas / NumPy / Elasticsearch / Kibana — cleaning, indexing, transforming and visualizing asset data.
LSTM multi-modal time-series model for climate-induced migration forecasting. Real-time data from 3+ weather APIs · +15% accuracy.
ML for precision water allocation in Indian agriculture — +35% water-use efficiency, automated expense categorization -50% manual entry.
LSTM forecasting on AWS Kinesis / S3 / SageMaker across 500+ houses & transformers — +25% grid resilience.
A beginner-to-mastery Python guide — kept as a public learning artifact.
your collaboration could live here. (yes, you, reading this.)
fetched from api.github.com/users/kevil21 when you opened this page.
42 contributions in the last year · view real graph on github
click any card to open the official verification page with credential ID prefilled.
cred id: 3d7ecea14b4342659eb4d68cdc9fbc37
verify ->cred id: b1f4b1e59f194f5da6872a5f1b79c824
verify ->cred id: 661B55BA40148A87
verify ->cred id: 141508FF683C2BFE
verify ->cred id: CBEF8794CADFA2F1
verify ->cred id: 505C219976CD2E5D
verify ->azure ai foundry · agentic workflows
verify ->
cred id: a2df505f-955c-41b6-a071-f21178c8a5bb
verify ->cred id: bbf6e80b-21cc-4306-9aea-77d03d0de28c
verify ->cred id: 09b81d9b-3aa0-4734-a85a-b9cd42c09e29
verify ->VIT University, Vellore · CGPA 8.95 / 10 · AIR 56 VITEEE
L.P. Savani International, Surat · GPA 8.9
L.P. Savani International, Surat · GPA 8