AI AgentsFull StackAIIn Progress

NirmanAI 2.0

A construction-planning platform combining a Gemini-powered conversational agent with deterministic engineering calculators, so answers are grounded either in retrieved IS-code knowledge or in verifiable computation — not just model output. Includes a bilingual (Hindi/English) chat assistant, a project planning wizard, a bill-of-quantities engine, a Vastu auditor, and a construction loan disbursement planner.

Source Code

Problem

Planning home construction in India means navigating BIS/IS codes, material quantity estimation, Vastu considerations, and loan disbursement schedules — usually scattered across contractors, spreadsheets, and guesswork, with no single grounded source of truth.

Solution

Built an agent architecture where a Gemini-powered chat assistant handles open-ended questions grounded in a curated IS-code knowledge base via RAG, while the parts of the problem that have exact answers — material takeoff, EMI schedules, Vastu scoring — are handled by deterministic calculators the agent can call, so the system never has to guess at something computable.

Architecture

01FastAPI backend exposing /chat, /report, and calculator endpoints (BOQ, Vastu audit, loan schedule)
02construction_agent.py orchestrates Gemini generation, RAG retrieval, and heuristic grounded synthesis
03rag_engine.py indexes an 18+ document Indian-construction knowledge base in ChromaDB, with a BM25 in-memory fallback
04construction_engine.py implements deterministic calculators for BOQ takeoff, Vastu evaluation, and loan scheduling
05React + Vite frontend with a per-tool page structure (chat, planner, estimator, Vastu auditor, loan planner, knowledge hub)

Implementation

▸Curated and structured the IS-code knowledge base (IS 456, IS 875, IS 1893, IS 13920, NBC 2016) for retrieval
▸Built the deterministic calculators first so the system has a reliable fallback path independent of any LLM call
▸Wrote an automated Pytest suite for the construction engine and agent logic
▸Implemented the Render free-tier keep-alive strategy (frontend wake banner + GitHub Actions ping) to manage cold starts

Challenges

▸Deciding which parts of a construction query should be answered by retrieval + generation versus deterministic computation, and keeping that boundary consistent
▸Supporting bilingual (Hindi/English) conversation without losing precision in technical, code-referenced answers

Result

A working full-stack agent platform with a passing automated test suite, combining grounded generation with deterministic engineering tools — actively being developed.

TECHNOLOGIES

PythonFastAPIReactGemini AIChromaDBRAG

KEY FEATURES

Bilingual (Hindi/English) AI chat assistant grounded in Indian construction (IS code) knowledge, with text-to-speech playback
Multi-step project master planner covering soil strata, Vastu, green amenities, and FAR bylaws
BOQ & cost takeoff engine computing exact material quantities (cement, steel, sand, bricks, tiles) with CSV export
Interactive Vastu Shastra auditor with an 8-direction room mapper and a live harmony score
Construction loan planner modeling a 7-stage disbursement schedule with pre-EMI and full-EMI calculations
Deterministic engineering fallback — core calculators work even without a Gemini API key

TECHNICAL HIGHLIGHTS

Construction agent combines Gemini generation with a RAG layer and deterministic heuristics for grounded synthesis, rather than relying on the LLM alone
RAG engine backed by ChromaDB with an in-memory BM25 fallback for semantic search over construction knowledge
Automated test suite (Pytest) covering the construction engine and agent logic
React 18 + Vite frontend rendering Markdown responses, Mermaid diagrams, and interactive SVG blueprints

MCPify

IntelliDocs

More projects

MCPify

Analyzes any AI agent's URL, detects its framework, and generates ready-to-use MCP configs — and exposes itself as an MCP server.

PythonFastAPIfastapi-mcpAPScheduler+1

In ProgressSource

IntelliDocs

A multi-user RAG platform: upload PDFs, ask questions, and get AI answers with page-level citations. Live and deployed.

PythonFlaskPostgreSQLJina AI Embeddings+3

LiveSource

AI-Powered Customer Support Assistant

A RAG-based support system that answers customer queries, classifies them, predicts priority, and scores its own confidence.

PythonFlaskPostgreSQLFAISS+3

LiveSource

NirmanAI 2.0