I'm a DevOps engineer with 3 years in cloud infrastructure. In my current role I'm the sole DevOps engineer on ciATHENA, an agentic AI analytics platform for life-sciences clients. I took it from its first environment to launch:
- 29 container apps × 3 environments on Azure Container Apps, 85+ Terraform resources per environment
- 10k+ requests a day from 250+ users
- Approval-gated releases on GitHub Actions and Azure DevOps, with revision-based auto-rollback
- GitHub OIDC instead of static cloud keys, Entra SSO, ISO 27001 evidence and SOC 2 preparation
- Azure OpenAI and AWS Bedrock in production, on regulated life-sciences data
Outside work I build open-source prototypes for problems I run into on cloud and AI platforms, write about them on Medium, and contribute fixes upstream.
👔 Hiring? The 30-second version
- 3 years in DevOps and cloud. Sole DevOps engineer on a production agentic AI platform, from first environment to launch.
- Azure + AWS in production: Container Apps, ECS Fargate, Terraform, GitHub Actions / Azure DevOps, OIDC, Entra SSO.
- AI infrastructure on regulated data: Azure OpenAI and AWS Bedrock, ISO 27001 evidence, SOC 2 preparation.
- Certified: AWS CloudOps Engineer – Associate, HashiCorp Terraform Associate (004).
- Proof beyond the day job: an installable Argo Rollouts plugin in Go, 18 benchmarked open-source prototypes, 20+ PRs merged upstream, 10 technical articles.
🛠️ Engineer? Start with these three repos
- policy-synapse: one policy, four enforcement points, and a differential tester that catches where they disagree.
- hephaestus-ledger: signing the image is not enough. This proves every live AI deployment part came from the approved release.
- chaos-witness: chaos tests aimed at the guarantees you promised, scored by how many of them a campaign actually checked.
Each repo has its architecture diagram, the benchmark and the limits in its README.
📖 Prefer reading? Start with these articles
| Project | What it does |
|---|---|
| rollouts-plugin-metric-sequential | Argo Rollouts metric plugin in Go. Aborts a canary when Prometheus error counts show a real regression, using a sequential test that stays valid however often Argo checks it. Worst false rollback 1.7%, against 30.6% for a z-test checked every interval; a 2x regression is caught in about 7 intervals. A kind end-to-end test in CI runs it inside the real Argo Rollouts controller. v0.1.0 ships with SHA-256 sums and SLSA build provenance. |
18 projects. Each one takes a single failure mode, builds a working prototype, and measures it against a baseline on a synthetic or simulated benchmark. Every README states the method, the numbers and the limits.
Security, supply chain and compliance
| Project | What it does |
|---|---|
| policy-synapse | Tests one policy across CI, Kubernetes admission, Terraform plan and runtime Rego. Caught 100% of 708 injected bugs; hand-written policies caught 56.9%. |
| hephaestus-ledger | Checks that every live AI deployment part (image, SBOM, Terraform state, model config, policy, identity) came from the approved release. 19/19 breaks caught; image signatures alone caught 1/19. |
| aegis-nexus | Narrows credentials at every hop when AI agents delegate to each other across clouds. Leakage after compromise: 242,782 → 7 weight-hours. |
| janus-gate | Finds which GitHub Actions identities a cloud OIDC trust policy really accepts (AWS, Entra ID, GCP) and fails CI on the extras. |
| chronos-keys | Plans secret rotation in a safe order across caches, restarts and scheduled jobs. |
| argus-evidence | Builds ISO 27001 / SOC 2 control evidence continuously from cloud config and logs, in a hash-chained bundle. |
| ariadne-thread | Traces sensitive fields through OpenTelemetry Collector pipelines to the backend that stores them. |
| penelope-loom | Gates AI-agent tool calls by side effect, and undoes them when a plan fails. |
| cassandra-upgrade | Scores dependency updates by whether your code reaches the changed functions and your tests cover them. |
Reliability and SRE
| Project | What it does |
|---|---|
| chaos-witness | Chaos testing aimed at explicit guarantees, not random faults. Found 7/7 planted weaknesses; a manual checklist found 4/7. |
| faultweaver | Lets an automated fix run only when independent telemetry backs the root cause. Unsafe actions: 29 → 11 across 240 incidents. |
| burn-sentry | SLO burn-rate alerts that stay trustworthy on low-traffic services. |
| orpheus-rollback | Decides if a release can really be rolled back across code, schema, config and feature flags. |
| canary-tribunal | Statistically valid promote / rollback calls for LLM features scored by noisy LLM judges. |
Platform, IaC and cost
| Project | What it does |
|---|---|
| terracog | Terraform drift triage: revert, codify, quarantine or escalate. Regret 0.59, against 2.75 for always-revert. |
| titan-fleet | One controller for LLM API routing and GPU autoscaling under SLO budgets. Beat the baseline in 23 of 25 scenarios. |
| midas-lens | Explains AI cost spikes by the deploy, prompt or config change that caused them. |
| carbon-oracle | Carbon-aware scheduling of AI jobs across regions, with a bound on missed deadlines. |
Also: gh-oidc-trust-fixtures (labelled Terraform fixtures for OIDC trust-policy scanners) · llm-gateway-quota-overshoot (reproducible harness for a LiteLLM budget-overshoot bug)
Every project ships its architecture as a diagram. Click one to open it, then drag to pan and scroll to zoom.
rollouts-plugin-metric-sequential: anytime-valid canary analysis inside Argo Rollouts
flowchart LR
P[Prometheus] -->|bad and total events,<br/>canary and baseline,<br/>since last measurement| R[Plugin Run]
S[Previous measurement metadata:<br/>test state + window end] --> R
R --> T[Sequential test<br/>mixture likelihood ratio<br/>H0: canary rate <= 1.1 x baseline]
T -->|e-value >= 1/alpha| F[Failed: Rollout aborts]
T -->|otherwise| OK[Successful: Rollout continues]
T --> S2[New measurement metadata:<br/>state, e-value, window counts]
aws-ecs-platform-blueprint: Terraform + ECS Fargate platform on AWS
flowchart TB
user([Client])
subgraph aws["AWS Region - ap-south-1"]
subgraph vpc["VPC 10.20.0.0/16"]
subgraph public["Public subnets - 2 AZs"]
alb["Application Load Balancer<br/>access logs to S3"]
nat["NAT Gateway"]
end
subgraph private["Private subnets - 2 AZs"]
task1["ECS Fargate task<br/>ARM64 - non-root"]
task2["ECS Fargate task"]
end
subgraph dbsub["Database subnets - 2 AZs, no internet route"]
rds[("RDS PostgreSQL 18.6<br/>encrypted - PITR")]
end
end
ecr[["ECR<br/>immutable tags"]]
secrets[["Secrets Manager<br/>RDS-managed password"]]
cw[["CloudWatch<br/>logs - metrics - alarms - dashboard"]]
sns[["SNS to email / Slack"]]
end
user -->|"80 / 443"| alb
alb -->|"8000"| task1
alb -->|"8000"| task2
task1 -->|"5432 TLS"| rds
task2 -->|"5432 TLS"| rds
task1 -.->|egress| nat
ecr -.->|image pull| task1
secrets -.->|"DB_USER / DB_PASSWORD"| task1
task1 -.->|"JSON logs + metrics"| cw
alb -.-> cw
rds -.-> cw
cw --> sns
policy-synapse: one policy intent, tested across four enforcement points
flowchart LR
I[intents.json<br/>governance intent IR] --> C[compiler]
C --> CI[CI Rego<br/>K8s manifest]
C --> AD[Gatekeeper<br/>AdmissionReview]
C --> TF[Terraform plan Rego<br/>plan JSON]
C --> RT[runtime Rego<br/>deploy request]
H[hand-written policies] -.-> D
I --> S[PSF boundary suite]
CI & AD & TF & RT --> D[differential engine<br/>OPA]
S --> D
D --> R[fingerprints, GCS,<br/>minimal counterexamples]
hephaestus-ledger: release provenance for AI deployments
flowchart LR
AR[Release root<br/>approver-signed:<br/>env, sequence, run, commit] --> RUN[CI run<br/>OIDC identity + commit]
RUN -->|SLSA provenance| I[Image digest]
RUN -->|CycloneDX claim| S[SBOM]
RUN -->|plan claim| TF[Terraform plan]
RUN -->|model claim| M[Model manifest + config hash]
RUN -->|policy claim| PB[Policy bundle + version]
TF -. pins digest .-> I
TF -. declares .-> WI[Workload identity + roles]
TF -. declares endpoint .-> M
TF -. must equal .-> ST[Applied state]
L[Live assembly<br/>from runtime observer] --> V{Reachable from<br/>the active root?}
AR --> V
V -->|every component| OK[exit 0, RGS 0]
V -->|gaps| GAP[per-component reasons, RGS, exit 1]
aegis-nexus: attenuating delegation chains for AI agents
flowchart LR
U[user consent] --> G[delegation gateway<br/>policy decision]
I1[aws.sim issuer] --> WP[WIT planner]
I2[azure.sim issuer] --> WR[WIT researcher]
I3[gcp.sim issuer] --> WC[WIT coder / linter]
WP --> G
G -->|hop 0: user entitlement, 1 h| P[planner]
P -->|hop 1: task caps, 10 min| R[researcher]
P -->|hop 1: task caps, 15 min| C[coder]
C -->|hop 2: narrower, 5 min| L[linter]
WR -.bound.-> R
WC -.bound.-> C
R & C & L -->|chain + PoP| T{tool verifier<br/>offline}
T -->|allow + audit path| X[search / docs / repo / mail / model / storage]
T -->|deny| D[reason]
janus-gate: who a CI-to-cloud OIDC trust policy really lets in
flowchart LR
TP[Trust configuration<br/>AWS trust policy / Entra FIC / GCP provider + bindings] --> E[Evaluator<br/>per-cloud condition semantics]
TP --> W
I[Declared intent<br/>repos, refs, environments, PRs] --> W[Witness claim space<br/>intent + policy constants +<br/>look-alikes + wildcard members]
CS[GitHub Actions claim model<br/>sub format, precedence, immutable IDs] --> W
W --> E
E --> S[Accepted tokens]
I --> D{Diff}
S --> D
D --> X[Excess identity classes<br/>+ witness subject + TES]
D --> G[Gated findings<br/>depend on GitHub environment rules]
D --> U[Locked-out intent]
X --> C[CI gate: exit 1]
chronos-keys: dependency-aware secret rotation
flowchart LR
INV[Declared consumer inventory<br/>refresh kind, TTL, delay bound,<br/>restart time, job schedule] --> G[Rotation Dependency Graph]
ST[Store bounds<br/>add propagation] --> G
G --> P[Staged planner<br/>add, promote, restart, wait, revoke]
P --> PLAN[Ordered plan + safe-by times]
PLAN --> GATE[Human approval / runbook]
P --> SIM[Rotation simulator<br/>versions, caches, restarts, jobs]
LOG[Access log<br/>usage guard, optional] -.-> P
FLEET[Synthetic fleets<br/>incl. shadow + stale consumers] --> SIM
SIM --> M[RBE + time-to-revoke<br/>failure taxonomy]
argus-evidence: continuous ISO 27001 / SOC 2 evidence
flowchart LR
CFG[Config history<br/>changes, snapshots,<br/>recorder heartbeats] --> TL[Control Operation Timeline<br/>hourly: verified / violated / unknown]
LOG[Activity log<br/>API calls, sign-ins,<br/>hourly delivery digests] --> TL
TIX[Ticket export] --> TL
CTL[Control specs<br/>Python properties] --> TL
TL --> ESS[Evidence Sufficiency Score]
TL --> GAP[Gap report]
TL --> BUN[Hash-chained bundle<br/>events + intervals]
BUN --> VER[verify: chain + recompute<br/>+ optional anchored head]
VER --> AUD[Auditor review]
ariadne-thread: sensitive-data lineage through telemetry pipelines
flowchart LR
SRC[Source declarations<br/>fields + labels per service] --> G[Collector graph<br/>receivers, pipelines,<br/>processors, connectors, exporters]
CFG[Collector config<br/>JSON-encoded] --> G
G --> TEP[TEP analyser<br/>label propagation on<br/>symbolic field templates]
TEP --> P[Exposure paths<br/>source, label, form,<br/>backend, region, retention, field]
CAN[Canary values<br/>3 formats per label] --> EMU[Pipeline emulator<br/>concrete values]
G --> EMU
EMU --> ST[Backend storage]
ST --> GT[Ground truth<br/>canary search: raw and SHA-1]
GT --> CMP{compare}
P --> CMP
penelope-loom: compensation-gated tool execution for AI agents
flowchart LR
AG[Scripted agent plan<br/>+ one injected fault] --> GATE{Effect gate}
REG[Tool registry<br/>effect class + compensation contract] --> GATE
GATE -->|read / reversible| EXE[Execute with<br/>idempotency key]
GATE -->|compensable + contract| EXE
GATE -->|compensable, no contract| BLK[Block]
GATE -->|irreversible| ESC[Escrow in ledger]
EXE --> LED[(Effect ledger)]
ESC --> LED
LED --> OK{Plan commits?}
OK -->|yes| APP[Human approves each<br/>held effect, then release]
OK -->|no / crash| UNW[Discard escrow,<br/>unwind in reverse,<br/>guarded compensations]
UNW --> HUM[Escalate what<br/>could not be undone]
cassandra-upgrade: break risk for dependency updates
flowchart LR
OLD[library vN<br/>parsed, never imported] --> DIFF[API diff<br/>removed, signature, default,<br/>body, transitive, changelog flag]
NEW[library vN+1<br/>parsed, never imported] --> DIFF
APP[client source] --> CG[call graph from the entry point<br/>+ argument shape of each call]
TEST[client tests on vN<br/>line tracer] --> COV[covered call sites]
DIFF --> R[reachable changed call sites<br/>+ call-site binding check]
CG --> R
R --> S[Break Risk Score<br/>noisy-OR over evidence]
COV --> S
CI[client tests on vN+1] --> D{red build, or<br/>score >= 0.5?}
S --> D
D -->|yes| B[block, with the call sites that explain why]
D -->|no| M[merge]
GT[ground truth: run the client workload<br/>on vN and on vN+1] -. evaluation only .-> E[benchmark]
chaos-witness: chaos campaigns aimed at explicit obligations
flowchart LR
C[contract.json<br/>10 obligations] --> G[campaign generator<br/>failure class -> proxy faults]
G --> P1[LLM proxy] & P2[tool proxy] & P3[identity proxy] & P4[store proxy]
W[client workload<br/>3 runs, 2 tenants] --> O[orchestrator under test]
O --> P1 & P2 & P3 & P4
P1 & P2 & P3 & P4 --> M[mock LLM / tool / IdP / store]
P1 & P2 & P3 & P4 -. request log .-> V[obligation checks]
M -. ledger, store, issued tokens .-> V
W -. status, latency .-> V
V --> A[verdicts, counterexamples, ACC]
faultweaver: evidence-bounded remediation
flowchart LR
S[Simulator: 17 services + 1 third-party API<br/>fault injection, blind spots,<br/>shared and faulty exporters] --> O[Observed telemetry<br/>metrics, logs, traces, change events]
O --> G[Hypothesis generator<br/>graph-constrained, deterministic]
G --> W[Diagnostic Witness Graph<br/>cause + propagation path + witnesses]
W --> C{DWC >= tau?}
C -->|no| P[Falsifying probe<br/>read-only, budget 3]
P --> O
C -->|yes| A[Remediation proposal]
C -->|no probe left| E[Escalate to human]
burn-sentry: uncertainty-aware SLO burn-rate alerting
flowchart LR
REQ[Request outcomes<br/>per minute] --> W[Workbook long windows 1 h / 6 h<br/>+ traffic-aware short window]
W --> POST[Beta-Binomial posterior<br/>per window]
POST --> D{P burn > k x budget<br/>>= c ?}
D -->|yes, short window agrees| PAGE[Page]
D -->|no, but P >= 0.2<br/>and posterior wide| PROBE[Synthetic probe session<br/>6 req/min for 10 min]
PROBE --> REQ
D -->|no| OK[No action]
orpheus-rollback: release reversibility analysis
flowchart LR
REL[Release manifest<br/>changes + declared inverses<br/>+ declared code reads/writes] --> C[Change classifier<br/>reversibility type]
C --> G[Release Reversibility Graph<br/>nodes: changes, old/new code<br/>edges: couplings]
G --> V{Verdict}
V -->|safe / safe with steps| PLAN[Ordered rollback plan]
V -->|unsafe| REF[Refuse: roll forward<br/>or escalate]
PLAN --> SIM[SQLite simulator<br/>traffic after every step]
ORA[Oracle: execute every<br/>plan from the catalogue] --> GT[Ground truth]
SIM --> EV[Errors / data-loss evidence]
GT --> EV
canary-tribunal: anytime-valid canary decisions for LLM features
flowchart LR
T[Paired traffic:<br/>each sampled request<br/>replayed on both versions] --> B[Baseline outputs]
T --> C[Canary outputs]
B --> J[LLM judge scores<br/>simulated: noise, position bias,<br/>verbosity bias]
C --> J
H[Human-labelled<br/>calibration set] --> N[Judge-bias interval<br/>E judge diff - human diff]
J --> CS[Anytime-valid confidence<br/>sequence on the paired difference]
N --> Q[Quality interval:<br/>CS shifted and widened]
CS --> Q
GR[Latency / cost / refusal<br/>confidence sequences] --> D{Deterministic rule}
Q --> D
D --> P[Promote / rollback / continue]
terracog: intent-aware Terraform drift triage
flowchart LR
TF[terraform show -json<br/>resource_drift] --> G[Intent Delta Graph]
EV[CloudTrail records<br/>principal, session, ticket] --> G
IT[ITSM export<br/>incident / change status] --> G
W[world.json<br/>service graph, SLO burn,<br/>policies, directory] --> G
G --> I[intent + confidence]
G --> R[REVERT end state:<br/>policy on IaC value,<br/>capacity or traffic lost]
G --> C[CODIFY end state:<br/>policy on live value,<br/>capacity cut kept]
I --> S[expected harm<br/>per action]
R --> S
C --> S
S --> O[REVERT / CODIFY /<br/>QUARANTINE / HUMAN_REVIEW<br/>+ Markdown evidence]
titan-fleet: LLM API routing + GPU autoscaling under SLO budgets
flowchart LR
R[requests<br/>interactive / batch / background] --> P[slow planner<br/>Holt forecast, one cold start ahead]
R --> Q[fast router<br/>marginal $ + budget-priced P miss]
B[error-budget tracker<br/>per class] --> Q
P -->|replicas| S[self-hosted pool<br/>queue, cold start, $/GPU-h]
H[health: 5xx / 429 cooldowns,<br/>metered quota, queue depth] --> Q
Q --> A[apiA-east / apiA-west<br/>Azure-OpenAI-like]
Q --> BB[apiB-east<br/>Bedrock-like]
Q --> S
Q --> M[Prometheus-style counters]
midas-lens: change-attributed cost anomalies for AI workloads
flowchart LR
USE[Usage logs<br/>requests, tokens<br/>per route x tenant x model] --> SER[Factor series<br/>volume, mix,<br/>tokens in/out, price]
PRICE[Price table] --> SER
SER --> CP[CUSUM change points<br/>per factor x scope]
CP --> MAT[Materiality check<br/>LMDI cost contribution]
EVT[Change log<br/>deploy, prompt, config,<br/>routing, tenant, price] --> ATT[Alignment<br/>scope + type + time]
MAT --> ATT
ATT --> REP[Ranked causes<br/>USD/h each, or UNLOGGED]
carbon-oracle: risk-bounded carbon-aware job scheduling
flowchart LR
J[synthetic AI jobs<br/>class, runtime, deadline,<br/>residency zone] --> F[feasible windows<br/>allowed regions x start slots]
C[GB regional carbon<br/>forecast] --> O[RGW optimizer<br/>min carbon + cost]
H[history of queue-wait and<br/>runtime errors] --> B[chance constraint<br/>P miss <= eps]
F --> O
B --> O
O --> P[placement: region, submit time]
P --> S[simulator: realised carbon,<br/>queue wait, runtime]
S --> R[Green Regret vs<br/>hindsight oracle]
20+ pull requests merged into other people's projects. See every merged PR →
| Project | Contribution |
|---|---|
| scylladb/python-driver | Set the Python 3 language level for the Cython build |
| Consolonia/Consolonia | CI: NativeAOT smoke test |
| Kareadita/Kavita | Fix seams between images in the webtoon reader |
| traceopt-ai/traceml | Hide the empty step-memory card on CPU-only runs |
| Soneso/stellar-agent-wallet | Silence intentional ShellCheck warnings on macro metavariables |
| Stellar-Developer-Observatory | CI: dependency and license audit |
| flisk-sh/citable | Return 404 for static-file paths |
| repo-engineering-skills | Verify environment defaults in docs (+1 follow-up) |
| hettiger/clg | Configurable release timezone |
| adjective-rob/snap | Windows setup script |
| Project | What it is |
|---|---|
| aws-ecs-platform-blueprint | Terraform + ECS Fargate on AWS: 3-tier VPC, RDS PostgreSQL, gated GitHub Actions delivery, Prometheus/Grafana and CloudWatch |
| aws-ecs-terraform | Internet → ALB → ECS Fargate → RDS, plus local Postgres with migrations, seed data and self-verifying backup/restore |
| ansible-linux-lab | Ansible done properly: inventory as code, idempotent roles, validated handlers, staged rollouts, health checks |
| wisecow-kubernetes-tls | A legacy Bash service containerised and run on Kubernetes behind NGINX Ingress with cert-manager TLS |
| flask-counter-ecr-pipeline | Source to resource-capped container: Docker, SonarQube, GitHub Actions to Amazon ECR |
| strapi-ephemeral-ec2-environments | Disposable environments on EC2: Terraform plus on-demand deploy and destroy workflows |
| azure-vm-health-check | Bash sweep of VMs across Azure subscriptions, CPU/memory/disk as CSV with threshold alerts |
| selfhosting-reference-library | Compose stacks, Ansible playbooks and reverse-proxy configs for ~90 self-hosted services |
I write up each project on Medium: the problem, what I built, the numbers and the limits.
- Your Chaos Test Passed. Which Guarantee Did It Actually Check?
- The Linter Should Not Hold the Keys to the Kingdom
- Right Service, Wrong Fix
- When Terraform and the Cloud Disagree, Which One Is Right?
- Your LLM Gateway and Your GPU Autoscaler Don't Share a Budget
- Your Image Is Signed. Is Your AI Deployment?
- One Rule, Four Policies: Testing Policy-as-Code Across Enforcement Surfaces
- AI Agents Are Writing Your Infrastructure Code. Is Anyone Governing It?
- Forty Minutes Was Enough
- Deploy Without Fear
Cloud: Azure (Container Apps, App Service, Front Door, APIM, Key Vault, Logic Apps, Azure OpenAI, Cosmos DB) · AWS (ECS Fargate, EKS, ECR, Lambda, EventBridge, S3, RDS, CloudFront, Bedrock) · GCP
Infrastructure as code and containers: Terraform (modular, multi-environment) · Ansible · Docker · Kubernetes · Helm · NGINX Ingress · cert-manager
CI/CD and GitOps: GitHub Actions · Azure DevOps · GitLab CI · Jenkins · Argo CD · GitHub OIDC
DevSecOps: Trivy · SonarQube · Gitleaks · Bandit · Safety · Snyk · OPA / Rego · Checkov
Observability: Prometheus · Grafana · Log Analytics · CloudWatch · CloudTrail · Datadog · ELK
Languages: Python · Bash · SQL · JavaScript
MCA, Manipal University Jaipur (CGPA 9.01) · BCA, Chandigarh University · Employee of the Quarter, Q1 2025
Former GDSC Cloud Lead: taught cloud to 500+ students through hands-on workshops
Widgets: GitSkins (Midnight) · capsule-render · readme-typing-svg · github-readme-stats · streak-stats · Platane/snk


