Nitesh Saini

// professional experience

1. Lead, Technology & AI · Bharti Institute of Public Policy, ISB

   Oct 2025 – Present

   Team & AI Infrastructure

   • Lead a team of 8+ GIS, software and AI engineers, owning the stack from field capture to models and deployment
   • Run self-hosted GPU servers (vLLM on 2× A30) for model training and on-premises Qwen2.5-VL 7B inference
   • Own the ODK survey infrastructure — XLSForm design, pyxform validation, and ODK Central → PostgreSQL → S3 pipelines — behind field data collection for multiple state government programmes

   Agentic AI & GraphRAG

   • Multi-Agent Governance Engine: architected a 4-agent system converting multilingual community testimony into structured governance metrics, using cascaded retrieval, drift monitoring and human-in-the-loop review
   • Human Development Analysis: GraphRAG system over 6 volumes / 102 chapters answering with page-level citations and drafting 9-section policy briefs, with 2-tier model routing and a reviewer agent validating each section

   nCount — Geospatial Intelligence Platform

   • Built an offline-first Android + web platform with 5K+ users, collecting 500K+ field data points across 7+ states
   • Designed GPS boundary mapping with Kalman filtering, geofenced loop closure, pause/resume and crash-safe autocapture, replacing the ODK geoshape widget
   • Manage the platform's web dashboards and inventory mapping pipelines, fusing Sentinel-1 SAR and Sentinel-2 imagery with terrain covariates into species distribution layers

2. AI Systems Engineer · Bharti Institute of Public Policy, ISB

   Apr 2025 – Oct 2025

   Computer Vision, Agentic AI & Voice AI

   • MonoDBH: tree diameter from a single RGB frame using YOLOv8s-seg at 0.953 mask mAP@0.5 with Grounding DINO + SAM 2 fallback; released a 5.2 GB open dataset
   • Food System Visualization Engine: shipped a LangGraph multi-agent system integrating 4 programme sectors into one platform with 4 AI assistants for policy, conversational GIS, natural-language data queries and reports
   • Van Saathi: multilingual voice pipeline (Whisper, Sarvam STT/TTS, pyannote diarization of up to 20 speakers) turning gram sabha recordings and boundary maps into LLM-drafted forest plans for 548 villages

   Research & Field Platforms

   • Co-authored BarkVisionAI in Scientific Data (Nature, 2026): a 156,001-image bark benchmark covering 13 tree species
   • Developed a voice-based key-informant interview agent with Sarvam STT/TTS, a Flask proxy layer, and RAG over pgvector
   • Built the CommFor Android field app (ODK Collect fork) with a custom login flow and OkHttp/Retrofit API integration
   • Designed a STAC catalogue ingestion pipeline (stac.forestgovernance.in) for spatial and non-spatial research datasets

3. Full Stack Developer · Bharti Institute of Public Policy, ISB

   Jun 2023 – Apr 2025
   • RatiFi: built an offline-capable Forest Rights Act claim filing and tracking system, live as JharFRA in Jharkhand, field-trialled with Andhra Pradesh's State FRA Cell
   • ArthaSakhi: built an Android app digitising forest-produce trade for women-led producer companies, with multi-step KYC, per-produce indents, blockchain-recorded transactions and QR-verified printed receipts (on Google Play)
   • Shipped field-facing web tools, including a bamboo-pilot monitoring dashboard (React, Leaflet, AWS Amplify) for Malkangiri CFR areas and a multi-cohort TBL simulation platform (Firebase Firestore) for ISB
   • Delivered CFR boundary-digitisation tooling (OpenCV/Python) and GLC30 land-cover analysis for community forest rights research
   • Managed core AWS infrastructure for field data collection, including ODK Central storage and data pipelines

// selected projects

• nCountGeospatial intelligenceLaunch ↗

  Geospatial data platform: an offline-first Android capture app and a web platform, running the inventory and boundary mapping pipelines.

  500K+ Field data points7+ States deployed

• MonoDBHComputer visionDataset ↗

  Metric tree diameter from a single RGB frame — trunk segmentation plus pinhole projection, no reference object or specialised hardware.

  0.953 Mask mAP50130 FPS Inference

• Food System Visualization EngineAgentic AILaunch ↗

  One analytical system for agriculture, nutrition, rural development, and livelihood programmes — explored through four AI assistants for policy, maps, data, and reports.

  4 AI assistants4 → 1 Programme sectors unified

• Human Development Analysis PlatformRetrieval-augmented architecturesLaunch ↗

  Citation-grounded RAG over 102 chapters of published research — page-level sourced answers and long-form briefs checked by an independent reviewer agent.

  102 Chapters indexed2-tier Model routing

more projects

• Field Image Cleaning Pipeline: Two-phase pipeline for nCount field imagery: computer-vision deduplication followed by multi-LLM ensemble consensus to flag unusable images; benchmarked against manual review
• LMS AI Studio: Offline-first learning platform (PWA) built for people where connectivity ends, pairing an AI learning guide with AI-assisted course authoring