Empowering 1M+ Indian farmers with AI diagnostics, hyperlocal weather, digital marketplace, and expert consultations.

What it is: Production SaaS platform for Indian smallholder farmers — crop disease detection via vision AI, real-time mandi prices, expert consultations, and a full digital marketplace.

Architecture highlights:

• 28 Spring Boot microservices (100% parity migration from Node.js monolith) — auth, marketplace, crop, weather, diagnosis, notifications, payments, and more
• 162 AI agents defined as Claude Code subagents — engineering, QA, product, DevOps, security, ML, marketing, executive roles — orchestrated via CrewAI crews and a Virtual Software Company pattern
• Claude Code deeply integrated — 162 .claude/agents/ definitions with YAML frontmatter, hooks (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop), GPS MCP server for symbol graph indexing, Render MCP for deployments
• Multi-model AI stack: claude-sonnet-4-6 (default), claude-opus-4-8 (high-reasoning CFO/CTO roles), claude-haiku-4-5 (lightweight ops), Cerebras (inference), Groq (audio transcription), Gemini
• LLaVA vision model — local CPU inference for crop disease detection across 38 PlantVillage classes, with 2000+ lines of India-specific treatment database (Hindi/English/Telugu/Marathi)
• Multi-tenant SaaS — PostgreSQL schema isolation per tenant, X-Tenant-Id header enforcement, row-level security
• Observability: ELK Stack, Prometheus + Grafana, Spring Sleuth + Zipkin, Sentry
• Event streaming: Kafka 3.7, Redis 7, PostgreSQL 14 + pgvector

Agent-driven SDLC — 110+ workflows covering the full engineering lifecycle:

Workflow runner: single ./workflow-runner.sh WF-N TASK-ID dispatches full agent chain, logs every step to shared GitHub issue as paper trail.

Stack: Java 17 / Spring Boot 3.4 · Node.js 18 / Express 5 · Flutter 3 · React 19 + Vite · Python / FastAPI · PostgreSQL · Redis · Kafka · Docker / Kubernetes · Render · Vercel

The full engineering lifecycle automated with 162 Claude Code subagents and 110+ workflows.

• 18-phase WF-SDLC pipeline — PRODUCT → DESIGN → TRIAGE → ENGINEER → SECURITY → BUILD-GATE → QA → REVIEW (5-role fan-out) → PACKAGE → CHANGE-REQUEST → CAB → DEPLOY → VERIFY → ROLLBACK → HYPERCARE → DOCS → DIFF-VERIFY → CLOSE → FINALIZE; gates block the deploy boundary (build + QA + security + review + CAB must all pass) and VERIFY triggers auto-rollback on canary failure
• 162 role definitions in .claude/agents/ (engineering, QA, product, DevOps, security, ML, marketing, executive) with YAML frontmatter, hooks (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop)
• Self-improving loop: nightly harvest of GPS corrections → triage → prompt mutation below 7.5 threshold → commit; GEPA-style eval loop scores agents across 5 dimensions
• Runner: single ./workflow-runner.sh WF-N TASK-ID dispatches the full chain and logs every step to a shared GitHub issue as paper trail
• Also practicing for the Claude Code Architect Certification — claude_course (prompting, tool use, evals, extended thinking, RAG, vector DBs)

Parallel Python implementation of the 39 canonical multi-agent workflows, built on langchain + langgraph.

• Same orchestration, outside the Claude Code harness — runs from cron, CI, or any Python environment
• 18-phase WF-SDLC pipeline ported into LangGraph with tool role-layer strategy (Spec Kit = Architect, Beads = tracker, Gas Town = EM, dev backends)
• Zero-credential execution — agent calls run through CLI backends (opencode run free models, claude -p OAuth session), no ANTHROPIC_API_KEY needed
• Schema-first contracts (CS-1…CS-10), diff-first edit mode, mechanical mode to cut 50–90% of LLM calls, anti-pattern catalog drawn from production incidents

The AI review loop doesn't just comment — it applies its own fixes.

• AI-powered PR code review in CI routes each finding to the specialist agent who owns that codebase
• Review loop applies its own fixes behind a label — human approves the label, the loop fixes and re-verifies
• Review gate blocks the merge instead of narrating it — quality enforced at the PR level, with P0/P1/P2/P3 defect triage and parallel fan-out fix dispatch per defect group

Capstone for the AI/ML certification track at Masai.

• Module 1: full data pipeline — scrape → clean → convert → store → query over books.toscrape.com (requests + BeautifulSoup + pandas + SQLite)
• 174 books across 10 categories, SQL/pandas query equivalence checks, median-imputation cleaning policy, GBP→INR conversion
• Support assistant (RAG-based) and data-pipeline analytics modules as the course progresses

• 🔭 Kheti Sahayak — scaling to 1M+ farmers, completing microservices migration, adding pgvector RAG
• 🤖 Claude SDLC — agent-driven SDLC: 162 subagents, 110+ workflows, self-improving eval loop; studying for Claude Code Architect Certification
• 🦜 LangChain SDLC — porting the agentic SDLC to LangGraph (39 workflows), zero-credential agent backends
• 🔄 PR auto-fix — CI review loop that applies its own fixes behind a label and gates merges
• 🎓 Masai AI/ML certification — capstone data pipeline + RAG support assistant
• 🌱 Spring Boot 3.4 + Kafka microservices architecture at production scale

Java · Spring Boot · Node.js · Microservices · Claude Code & Agentic AI · LangChain / LangGraph · Playwright automation · Multi-tenant SaaS · PostgreSQL · Kafka · Docker/Kubernetes · Data Pipelines · ML

Email: p.ponali@yahoo.com
Blog: prakashponali.wordpress.com