Architecting Out Loud

Anmol Yadav

+971-586978210·anmol1696@gmail.com·github.com/Anmol1696·anmolbuilds.com

Staff-level infrastructure engineer and technical co-founder with 7+ years building Kubernetes-native platforms, AI agentic infrastructure, distributed systems, authorization systems, developer tools, and ML/data infrastructure. Built Constructive's multi-tenant Postgres platform, created Starship for reproducible dev/test environments, scaled petabyte geospatial ML pipelines at Toyota Woven Planet, and cut authorization latency 50x at Rakuten. Current work applies that platform background to AI agent runtimes, eval-gated automation, and self-hosted inference.

Download PDF

Experience

Constructive

Founding Engineer, later Co-Founder & Head of Infrastructure · Remote, Dubai, UAE

Oct 2023 – Present

AI Agent Infrastructure

  • Built an internal autonomous engineering agent platform on self-hosted Kubernetes for issue triage, PR review, CI debugging, and cross-repo failure analysis; reduced manual project-tracking work by auto-generating issues from meetings and Slack.
  • Operated self-hosted vLLM inference behind OpenAI-compatible APIs with long-context serving, tool calling, and eval gates for internal agents.
  • Built the eval-pipeline harness and autonomous AI QA environment that exercise the database cluster and services like real users before promotion.
  • Built Go runtime sidecars for agents, including a multi-provider OAuth PKCE credential manager and a workspace introspection HTTP server.

Platform & Developer Infrastructure

  • Led the platform shift from chain tooling to a multi-tenant Postgres developer platform, then built that platform across 134+ packages and 3 monorepos, enforcing tenant isolation with PostgreSQL row-level security and JWT claims.
  • Reduced recursive authorization checks from exponential traversal to logarithmic lookups by building materialized authorization tables across the app/org/group hierarchy.
  • Designed a bitstring permission model for constant-time authorization checks by materializing multi-source permission bundles into one actor/entity bitmask.
  • Reduced database-package drift by building pgpm, an npm-style SQL package manager with dependency resolution, deterministic migrations, and content-addressed artifacts.
  • Built 14+ GraphQL plugins covering unified search, full-text/vector/trigram search, geospatial queries, file uploads, and retrieval-backed application workflows.
  • Designed composable authentication across password, OAuth/OIDC, magic link, OTP, WebAuthn/passkeys, and TOTP 2FA.
  • Enabled database-backed apps to run distributed workloads by building constructive-functions, a modular application-functions framework hosted as Knative services and managed by the platform database operator.
  • Built a cross-repo CI/CD staging system that runs full-system E2E in GitHub Actions and gates GKE dev-cluster promotion; 175+ cycles, 67% of PRs auto-merged.
  • Built KubernetesJS, a zero-dependency TypeScript Kubernetes SDK generated from the Kubernetes OpenAPI spec.
  • Maintains open-source Postgres tooling including pgsql-parser, libpg-query-node, and postgres-ast-deparser.

Starship

Creator & Lead · Open Source

Feb 2023 – Present
  • Created Starship, an open-source Kubernetes platform that spins up ephemeral Cosmos, Ethereum, and Solana testnets with IBC relayers from declarative config; primary contributor with about 672 commits.
  • Grew Starship into an E2E testing environment used by Cosmos teams, with 100+ users, 10,000+ downloads, about $225k in Osmosis/Neutron development grants, plus $500k from ICF (Interchain Foundation).
  • Integrated Starship with cosmos-sdk, mesh-security, Agoric chain-sdk, Skip APIs, Interchain Security, and Neutron, leading mesh-security integration testing and simulation while coordinating across Confio, Cosmology, DAO DAO, and Osmosis.
  • Designed init-container coordination for validator bootstrapping, exposer sidecars over gRPC/HTTP, ConfigMap-backed discovery, and a chain-registry-compatible service registry.
  • Built schema-first deep-merge configuration and generated TypeScript types across 30+ chain types with incompatible runtime shapes, porting the client surface from Go to fully TypeScript workspaces.

Persistence / Audit.one

Tech Lead · Remote, Dubai, UAE

Mar 2022 – Feb 2023
  • Led a core team of 4 and owned code-review and CI standards for a 15+ engineer org.
  • Ran 20+ validator nodes across Cosmos networks and cut cloud costs 50%+ in 2 months by rebuilding validator infrastructure.
  • Maintained and upgraded the core-1 Cosmos SDK chain from version 2 to 6 over 6 months while coordinating with validator operators, building internal tooling, CI/CD, and release automation for validator and chain operations.
  • Prototyped Kubernetes-based multi-chain devnet tooling for CI that later became Starship.

Woven Planet (Toyota)

Software Engineer to MLOps Engineer · Tokyo, Japan

Aug 2020 – Mar 2022
  • Owned the entire Pachyderm + Kubernetes pipeline (orchestration, data-transfer layer, and scaling architecture) that turned sub-meter satellite imagery into HD maps for autonomous driving across Japan and the USA.
  • Scaled the geospatial ML pipeline to petabyte-scale data across 88 geographic regions using S3, PostGIS, Elasticsearch, H3 hex partitioning, and Detectron2 inference, splitting stages into shared-global and per-region parallelism.
  • Built a schema-first per-region config system (DeepChainMap) so all 88 regions inherited global defaults and overrode only deltas, eliminating large-scale YAML duplication.
  • Rewrote the pipeline controller onto the Pachyderm Python SDK with autoscaling and pod-patching, and added SHA-256 content dedup to make ingestion idempotent and safe under reprocessing.
  • Built a 3-input Pachyderm join for lane-network generation (image tiles, segmentation masks, geo-metadata) plus an ICP-based validation stage measuring lane-feature lateral error against PostGIS ground truth.
  • Maintained Inadatron, the team's PyTorch/Detectron2 perception platform, as its heaviest contributor outside the core team; built the training pipeline scaling onto 4x DGX-2, driven by GitHub Actions CI/CD instead of manual training triggers.
  • Cut CI cost and feedback time by replacing always-on EC2 GPU runners with autoscaled Kubernetes jobs, and stratified a 280+ test suite into tiered unit/multi-GPU/nightly stages mapped to compute classes.
  • Owned the CI/CD and automation around ONNX/TensorRT edge export (W&B checkpoints to per-camera TensorRT builds) with Hydra config and Weights & Biases tracking, plus data cleanup/sanitation gates.
  • Built Computron from scratch, a distributed file-compute framework over S3 + SQS with queue-driven autoscaling and pluggable gRPC compute modules processing millions of files.

Rakuten

Cloud Platform, Software Engineer to Cloud Architect · Tokyo, Japan

Oct 2018 – Aug 2020
  • Cut authorization latency 50x by replacing Keycloak with Madara, a Go + Dgraph graph-based IAM system that became the source of truth for access, billing, and resource definitions across the OneCloud private cloud (10 data centers, in production years after I left).
  • Modeled authorization as two interlocking graphs (policy and component) with dual-grain tenant- and resource-level access from one polymorphic role-binding, plus a unified hierarchical resource-naming scheme (RNS).
  • Achieved near-zero-latency authz by compiling the graphs into pre-computed OPA policy bundles served to sidecars, moving decisions to the edge with no runtime network calls (ETag/304-cached).
  • Made the system BCP-compliant and multi-region (Kubernetes/Kustomize, Dgraph over mTLS gRPC) with best-effort eventually-consistent reads and a linearizable escape hatch.
  • Kept identity in sync with the IdP through a Kafka + CloudEvents backbone and a companion consumer reconciling desired state into Keycloak.
  • Rewrote a prior JanusGraph/Python IAM into Go/Dgraph; helped design OneCloud's Management API gateway and built Automation-as-a-Service with Terraform and Habitat.
  • Ran internal talks on cloud design patterns adopted across OneCloud teams.

Projects

Minato

Go, Kubernetes, Python/FastAPI, Helm

  • Hybrid agentic workflow engine that runs agent workloads across environments: local, on-prem, and cloud. A cloud-native harness for provisioning, orchestrating, and isolating long-running agents.
  • Kubernetes-native control plane: custom CRDs (Tenant, Workspace) with a kubebuilder operator that provisions isolated per-user namespaces with NetworkPolicy, ResourceQuota, and AES-encrypted credential injection.
  • Sidecar-based skills architecture where agents self-manage K8s resources via MCP tools. Domain example: a radiology DICOM service for CT/MRI analysis with multi-window rendering and session-based radiologist feedback loops, deployed via Helm with GPU support.

Base Chain Games

Solidity, Ethereum/Base

  • Built and deployed on-chain games on Base using custody and distribution smart contracts.

Focus Analytics

KNN, SVM

  • WiFi indoor positioning (KNN + SVM); improved accuracy 15% via a stateful predictor.

Niki.ai

Java, NLP

  • NLP chat domain identification (hidden Markov, bag-of-words) and intent classification in Java.

AES Cipher Modification

Cython, C++

  • Irreducible polynomial with an extra key; NIST-validated, 25x Cython/C++ speedup.

Skills

Languages
GoPythonTypeScript/JavaScriptSQLBashC/C++
Infrastructure
KubernetesHelmkubebuilder operatorsDockerTerraformGitHub ActionsGKEEKSKnative
Data & Storage
PostgreSQLPostgreSQL RLSpgvectorpg_trgmPostGISDgraphElasticsearchCassandraKafkaS3/SQSH3 partitioning
Authz & Security
graph-based IAMOpen Policy AgentRBAC/ABACOAuth/OIDCWebAuthn/passkeysJWT
AI/ML & Agents
vLLMself-hosted inferenceagent runtimeseval pipelinesautonomous QAMLOpsPyTorchDetectron2ONNX/TensorRTDGX/on-prem training
APIs & Interfaces
gRPCGraphQLPostGraphileCloudEvents
Observability
GrafanaELK

Education

Indian Institute of Technology Bombay (IIT Bombay)

B.Tech in Aerospace Engineering

2014 – 2018

Research

Cryptanalysis using Boolean Satisfiability Solver

Guide: Prof. Virendra Sule, Electrical Department, IIT Bombay · eprint.iacr.org/2017/1141

Dec 2016 – Sept 2018
  • Co-authored a research paper introducing a new SAT solving method.
  • Solved a 90-variable biological gene regulatory problem using the SAT solver to obtain all fixed points.
  • Developed a computational algorithm for formation of equations and symbolic operations for Boolean functions.
  • Recovered an 80-bit Bivium key within 48 hours using only 90 equations.