Mahesh Moholkar

An incident brain: when an alert fires, Claude agents walk a graph of services, deploys, commits, people and past incidents to find the likely root-cause change, page the owner, and cite the last time it happened.

• Multi-Agent Root Cause: Built a FastAPI agent pipeline (Triage, Investigator, Historian, Responder, Scribe) that computes blast radius, ranks suspect deploys and config changes, and takes the cheapest path to them through the services that are sick right now.
• Graph Memory: FalkorDB is the only datastore: company graph, 90 days of history, agent memory and handoffs; similar past incidents come from local vector search (fastembed) reranked by what the graph knows, with the runbook that worked most often.
• War Room & Company Brain: Next.js war room streams every tool call, decision and handoff live over the graph canvas; agents page on-call on Discord, open GitHub issues with the suspect commit, and write postmortems that re-rank runbooks. Ask “what breaks if checkout-api goes down?” and get an answer with evidence nodes linked.

Python / FastAPI / Claude (Haiku + Sonnet) / FalkorDB / fastembed / Next.js / TypeScript / TailwindCSS / Docker

AI customer support you embed on a website: a chat and voice widget answered by an agent that searches the business's knowledge base and hands off to a human operator when needed.

• Grounded Agent: PydanticAI agents on FastAPI answer from a per-organization knowledge base (PDF, Word, Markdown, HTML, OCR for scanned files) using pgvector search, streaming replies and escalating to a person on request.
• Real-Time Voice: Pipecat voice bot with Silero VAD, Smart Turn v3, Whisper and Kokoro running locally on MLX, with live captions in the widget.
• Operator Inbox: Next.js dashboard where operators see every chat and call transcript, take over from the AI, hand back and resolve; Better Auth organizations with JWTs verified by FastAPI.

Python / FastAPI / PydanticAI / SQLAlchemy / PostgreSQL (pgvector) / Ollama / Pipecat / Whisper / Next.js / shadcn/ui / Better-Auth

Designed, built and run a live multi-tenant GST ERP for Indian SMEs end to end: React web app, Next.js storefront, Expo Android app and .NET API.

• Transactional Posting: Sales and purchase documents run as state machines (draft → posted → cancelled); posting moves stock, books double-entry ledger entries and updates balances in one transaction, and cancelling reverses it.
• GST Compliance: CGST/SGST/IGST split by place of supply, tax snapshotted when a document posts, GSTR-1/3B generated from posted documents, and filed periods locked against edits, cancels and back-dated posts.
• Tenant Isolation: Every tenant-owned table is scoped to the signed-in business by a global query filter in the data layer, so a missed WHERE clause can't leak another business's data.
• Contract-First Clients: One OpenAPI contract generates the types for the React web app, Next.js storefront and Expo mobile app, so an API change breaks the build instead of production.
• Inventory: Batch and expiry stock with first-expiry-first-out picking, multi-location transfers and moving-average valuation.
• AI: Claude reads a supplier's bill photo into a draft purchase bill with structured output; the mobile app finds products by photo using Cohere image embeddings in pgvector.
• Integrations: Storage, email, WhatsApp, push and the LLM each sit behind an adapter with a no-op fallback, so the system runs and is tested without any of them.

Next.js / React / TypeScript / TanStack Router / Expo / PostgreSQL / pgvector / .NET 10 / EF Core / Hangfire / SignalR / Claude / Cohere / Docker / Railway