[{"content":"Overview Maintenance-Eye is a real-time AI copilot for physical infrastructure maintenance. Built for the Google Gemini Live Agent Challenge 2026, it lets a technician point a phone camera at equipment, ask by voice, review asset context, and confirm work-order actions through a camera-and-voice inspection flow.\nLinks GitHub Repository Devpost Demo Video Status Maintenance-Eye is a public portfolio case study backed by the Devpost demo, repository, architecture diagram, infrastructure files, API routes, and tests.\nDemo And Review Path The Devpost page is the primary public demo because it includes the project video. The repository documents local setup with a JSON-backed EAM fallback, so the data, API, and tool behavior can be inspected locally; live Gemini inspection and chat require Gemini credentials.\nProblem Transit and infrastructure maintenance work is physical, noisy, and time-sensitive. Technicians often need to inspect equipment, recall safety procedures, search maintenance history, and create work orders while their hands are already occupied by tools and safety gear. Traditional enterprise systems force a stop-and-type workflow that interrupts inspections and slows response time.\nWhat I Built A real-time multimodal frontend that streams camera frames and microphone audio from a phone-based PWA A FastAPI backend that manages WebSocket sessions, media flow, confirmation state, and operational APIs A Google ADK agent powered by Gemini 2.5 Flash Live API for native audio + vision reasoning A set of tool-using maintenance workflows covering asset lookup, knowledge retrieval, work-order actions, inspection history, and safety protocols A human-in-the-loop confirmation layer for critical actions such as creating or updating work orders Tech Stack AI runtime: Google ADK, Gemini 2.5 Flash Live API Backend: Python, FastAPI, WebSockets Data/artifacts: Firestore EAM data, JSON-backed local fallback, and optional best-effort GCS storage for frame, report, and work-order artifacts Infra: Cloud Run, Docker, Terraform, GitHub Actions Testing: unit, integration, API contract, security, performance, and Playwright E2E smoke tests Architecture The system uses a persistent bidirectional WebSocket to move video frames, audio, transcripts, confirmation cards, media cards, and work-order result messages between the phone client and backend. The agent calls domain-specific tools, grounds responses in maintenance data, and requires explicit confirmation before sensitive actions.\nEngineering Highlights Real-time multimodal interaction with audio in, audio out, and camera-driven reasoning Nine ADK tool bindings covering search, asset lookup, inspection history, knowledge retrieval, work-order management, safety protocols, report generation, and confirmation workflows Human-in-the-loop safety with confirmation cards for critical actions Operational deployment path with Docker, Cloud Run, Firestore, and Terraform Multi-layer test coverage across unit, integration, API contract, security, performance, and Playwright E2E smoke tests Why It Matters This project demonstrates my approach to applied AI: model integration, backend tool orchestration, deployment, test coverage, and human-in-the-loop safeguards around operational workflows.\n","permalink":"https://avisheksaha.com/projects/maintenance-eye/","summary":"\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003eMaintenance-Eye\u003c/strong\u003e is a real-time AI copilot for physical infrastructure maintenance. Built for the \u003cstrong\u003eGoogle Gemini Live Agent Challenge 2026\u003c/strong\u003e, it lets a technician point a phone camera at equipment, ask by voice, review asset context, and confirm work-order actions through a camera-and-voice inspection flow.\u003c/p\u003e\n\u003ch2 id=\"links\"\u003eLinks\u003c/h2\u003e\n\u003cul\u003e\n\u003cli\u003e\u003ca href=\"https://github.com/sahaavi/Maintenance-Eye\"\u003eGitHub Repository\u003c/a\u003e\u003c/li\u003e\n\u003cli\u003e\u003ca href=\"https://devpost.com/software/maintenance-eye\"\u003eDevpost Demo Video\u003c/a\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"status\"\u003eStatus\u003c/h2\u003e\n\u003cp\u003eMaintenance-Eye is a public portfolio case study backed by the Devpost demo, repository, architecture diagram, infrastructure files, API routes, and tests.\u003c/p\u003e","title":"Maintenance-Eye"},{"content":"Overview GovIntel is a local-first federal procurement intelligence system. It imports public USAspending contract awards, stores them in PostgreSQL, builds a Chroma-backed retrieval index, and generates citation-grounded market briefs through a FastAPI API and Streamlit UI.\nI built GovIntel as a reviewable RAG application with ingestion, indexing, retrieval, analytics, generation, citation checks, Dockerized local setup, CI, tests, and a walkthrough.\nLinks GitHub Repository UI Walkthrough Status GovIntel provides a reviewable end-to-end path: seed contract data, build the retrieval index, run the FastAPI service, open the Streamlit UI, ask a procurement question, and inspect cited award evidence.\nI focused this portfolio version on engineering completeness: local setup, retrieval, generation, citation checks, tests, and a walkthrough, rather than production hosting or benchmark claims.\nReview Path The walkthrough shows agency and NAICS inputs, a 1-10 year range control, a DHS cybersecurity question, a generated brief, and cited contract evidence. For hands-on review, the repository documents local setup through make db-up, make db-seed, make index, make run, and make ui.\nProblem Federal procurement data is useful but hard to scan quickly. The goal of GovIntel is to turn contract records into structured intelligence briefs: contractor patterns, agency spending questions, strategic implications from contractor rankings, quarterly spend trends, concentration, and cited award evidence.\nWhat I Built Async USAspending award ingestion with pagination and idempotent PostgreSQL upserts Typed Pydantic models for awards, analysis requests, contractor summaries, retrieved evidence, and generated briefs Chroma-backed vector indexing with sentence-transformer embeddings BM25 keyword retrieval, vector retrieval, hybrid merge, and cross-encoder reranking SQL analytics for top contractors, quarterly spend trends, and market concentration Versioned prompt templates and structured JSON generation Fail-closed citation validation before returning a brief FastAPI /api/v1/analyze endpoint with optional X-API-Key protection Streamlit UI for choosing filters, generating briefs, and inspecting cited contract evidence Docker Compose stack for PostgreSQL, the API, and the UI Tech Stack Backend: Python, FastAPI, Uvicorn, Pydantic v2 Data: PostgreSQL 16, SQLAlchemy asyncio, asyncpg Retrieval: ChromaDB, sentence-transformers, BM25, hybrid retrieval, cross-encoder reranking Generation: Gemini provider path, structured JSON prompts, optional Hugging Face provider path Frontend: Streamlit Operations: Docker Compose, GitHub Actions, Ruff, mypy, pytest, pytest-cov Optional extensions: Langfuse tracing, Pinecone mirroring, offline evaluation and QLoRA utilities Architecture The system flow is:\nPull bounded USAspending award data into normalized contract models. Persist contract rows in PostgreSQL. Build local retrieval indexes with Chroma vector search and BM25 keyword search. Merge and rerank retrieval candidates. Compute SQL analytics for contractor ranking, spend trend, and market concentration. Render a procurement-intelligence prompt with retrieved context and analytics. Generate a structured brief through the selected provider. Validate citations against retrieved contract evidence before returning the answer. Serve results through FastAPI and the Streamlit UI. Engineering Highlights End-to-end application path with ingestion, indexing, API, Streamlit UI, evidence inspection, and regression tests Guarded RAG design that rejects briefs whose citation list references award IDs outside the retrieved contract evidence Hybrid retrieval stack combining lexical recall, vector retrieval, deduplication, and reranking Structured analytics layer for contractor rankings, spend trends, and market concentration Provider boundaries that keep external LLM use explicit and configurable Quality gates through Ruff, strict mypy, and pytest with a 90% coverage threshold across API, retrieval, generation, ingestion, frontend, evaluation, observability, and training tests Evaluation and Quality GovIntel includes evaluation fixtures and an ablation harness; I use them as methodology evidence rather than headline quality claims.\nWhy It Matters GovIntel demonstrates my RAG and data-systems work: retrieval, structured generation, analytics, citation validation, API design, local deployment, and evaluation scaffolding over real procurement data.\n","permalink":"https://avisheksaha.com/projects/govtintel/","summary":"\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003eGovIntel\u003c/strong\u003e is a local-first federal procurement intelligence system. It imports public USAspending contract awards, stores them in PostgreSQL, builds a Chroma-backed retrieval index, and generates citation-grounded market briefs through a FastAPI API and Streamlit UI.\u003c/p\u003e\n\u003cp\u003eI built GovIntel as a reviewable RAG application with ingestion, indexing, retrieval, analytics, generation, citation checks, Dockerized local setup, CI, tests, and a walkthrough.\u003c/p\u003e\n\u003ch2 id=\"links\"\u003eLinks\u003c/h2\u003e\n\u003cul\u003e\n\u003cli\u003e\u003ca href=\"https://github.com/sahaavi/GovtIntel\"\u003eGitHub Repository\u003c/a\u003e\u003c/li\u003e\n\u003cli\u003e\u003ca href=\"https://youtu.be/-qqBq8bCaSg\"\u003eUI Walkthrough\u003c/a\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"status\"\u003eStatus\u003c/h2\u003e\n\u003cp\u003eGovIntel provides a reviewable end-to-end path: seed contract data, build the retrieval index, run the FastAPI service, open the Streamlit UI, ask a procurement question, and inspect cited award evidence.\u003c/p\u003e","title":"GovIntel"},{"content":"Overview I published this undergraduate applied-ML research on toddler autism-screening data in the International Journal of Information Technology and Computer Science (IJITCS). The work combines dataset analysis, a Tableau/Python dashboard, and a literature review of SVM/CNN classification approaches.\nI include this project as early published research experience; my current applied-AI engineering case studies are Maintenance-Eye and GovIntel.\nPublication Published in: International Journal of Information Technology and Computer Science (IJITCS), Vol. 14, No. 4\nApproach Data preparation - prepared Kaggle toddler autism-screening data for analysis and candidate classification workflows Modeling - reviewed SVM/CNN approaches for classifying ASD-trait screening labels Evaluation - summarized classification metrics reported in the paper and related literature Visualization - used Tableau to communicate dataset patterns and results Tech Stack Python Tableau SVM/CNN classification methods Key Takeaway This page treats the work as early published research and dashboard/data-analysis evidence; performance claims are limited to what the paper documents.\n","permalink":"https://avisheksaha.com/projects/autism-spectrum-disorder-prediction/","summary":"\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003eI published this undergraduate applied-ML research on toddler autism-screening data in the International Journal of Information Technology and Computer Science (IJITCS). The work combines dataset analysis, a Tableau/Python dashboard, and a literature review of SVM/CNN classification approaches.\u003c/p\u003e\n\u003cp\u003eI include this project as early published research experience; my current applied-AI engineering case studies are Maintenance-Eye and GovIntel.\u003c/p\u003e\n\u003ch2 id=\"publication\"\u003ePublication\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003ePublished in:\u003c/strong\u003e \u003ca href=\"https://www.mecs-press.org/ijitcs/ijitcs-v14-n4/v14n4-2.html\"\u003eInternational Journal of Information Technology and Computer Science (IJITCS), Vol. 14, No. 4\u003c/a\u003e\u003c/p\u003e","title":"Autism Screening Data Dashboard Research"},{"content":"Applied AI engineer with a data science and analytics background, 4+ years of Python/SQL experience, and hands-on work building LLM applications, AI agents, and agentic workflows.\nI build practical systems with OpenAI API, Anthropic API, Gemini API, Google ADK, RAG, tool calling, structured outputs, evals, Langfuse tracing, FastAPI, PostgreSQL, Docker, and cloud deployment patterns.\nI hold a Master of Data Science from UBC and a BSc in Computer Science from AIUB (CGPA: 3.91/4.0). My public case studies include Maintenance-Eye, a real-time multimodal maintenance copilot built with Google ADK, Gemini Live API, FastAPI, Firestore, and Cloud Run.\nCurrently, I\u0026rsquo;m a Data \u0026amp; Applied AI Analyst at BC Rapid Transit Company where I build incident, KPI, maintenance, and asset-performance data workflows while mapping operational workflows into AI-ready use cases for decision support and diagnostics. GovIntel is my federal-procurement RAG project, built around USAspending ingestion, hybrid retrieval, SQL analytics, citation validation, FastAPI, and Streamlit.\nWhat I Build Applied AI systems — multimodal agents, RAG workflows, tool use, and human-in-the-loop safeguards Python backends and data workflows — FastAPI services, ETL pipelines, structured analytics, and data modeling Operational AI systems — systems shaped by real maintenance, reporting, and decision-support workflows Professional Experience BC Rapid Transit Company — Data \u0026amp; Applied AI Analyst Dec 2024 - Present | Burnaby, BC\nBuild Python and SQL pipelines that consolidate incident, KPI, maintenance, asset-performance, and operational records into reliability-analysis datasets for reporting, natural-language querying, and AI-assisted decision support. Translate maintenance and reliability workflows into AI-ready use cases across triage support, asset-context retrieval, recurring-failure investigation, KPI interpretation, and human-reviewed recommendations. Prototype and support Microsoft/Azure automation workflows using Azure OpenAI, Azure AI Foundry, Copilot Studio, Power Platform, and tool-based interfaces for source-grounded operational analysis. Improved recurring asset-performance validation and reporting workflows, saving 30-40 minutes per day while increasing consistency of maintenance review processes. Brain Station 23 — Data Scientist, Financial Analytics \u0026amp; AI Workflows Aug 2023 - Nov 2024 | Dhaka, Bangladesh / Remote\nBuilt SQL, Python, and Power BI workflows for financial metrics, trend analysis, data consistency checks, recurring business reporting, and stakeholder-facing dashboards across multiple source systems. Improved reporting efficiency by 21%, data consistency by 17%, and query performance by 26% through SQL optimization, ETL workflows, indexing, partitioning, Python preprocessing, and validation checks. Created repeatable reporting and analysis workflows using SQL window functions, stored procedures, metric definitions, and data-quality checks to reduce manual review and improve analytical trust. Contributed to financial analytics and document-intelligence workflows using Vertex AI, Cloud Functions, BigQuery, Python, SQL, and Power BI. Statistics Canada — Data Scientist (Capstone) Apr 2023 - Jul 2023 | Ottawa, ON\nImplemented clustering algorithms (K-means, hierarchical) to segment the Proximity Measure Database for urban planning Conducted data profiling, exploration, and outlier detection in R Delivered actionable insights through cluster profiling for urban development policy recommendations Softology IT — Co-Founder Sep 2021 - Aug 2022 | Dhaka, Bangladesh\nCo-founded a software development firm delivering Android apps, e-commerce solutions, and UI/UX design Managed end-to-end project delivery across multiple concurrent client engagements Selected Projects Maintenance-Eye — Applied AI Project 2026 | Public project\nBuilt and deployed a real-time multimodal maintenance copilot using Google ADK, Gemini 2.5 Flash Live API, FastAPI, Firestore, and Cloud Run Implemented tool-using workflows for asset lookup, knowledge retrieval, work-order actions, and human-in-the-loop confirmation Added deployment and reliability support with Terraform, Docker, GitHub Actions, and a multi-layer test suite GovIntel — Federal Procurement Intelligence RAG System 2026 | Public project\nBuilt a local-first procurement intelligence RAG system over USAspending awards Implemented async ingestion, PostgreSQL persistence, Chroma indexing, BM25/vector retrieval, hybrid deduplication, cross-encoder reranking, SQL spend analytics, and fail-closed citation validation Added a FastAPI analysis endpoint, Streamlit UI, Docker Compose workflow, CI, and automated tests Education University of British Columbia — Master of Data Science (Sep 2022 - Jun 2023)\nAmerican International University-Bangladesh — BSc in Computer Science \u0026amp; Engineering, CGPA: 3.91/4.0 (Jan 2018 - Sep 2021)\nCertifications TensorFlow on Google Cloud | Google Cloud | 2024 Machine Learning Operations (MLOps) | Google Cloud | 2024 Machine Learning Specialization | Coursera | 2024 Skills AI Engineering and LLM Systems: Azure OpenAI, Azure AI Foundry, Anthropic API, Google Vertex AI, Gemini API, Google ADK, LangChain, LangGraph, LlamaIndex, LLM applications, AI agents, agentic workflows, RAG, tool calling, structured outputs, source grounding, human-in-the-loop workflows, prompt engineering\nAutomation, Evals, and Observability: AI automation, workflow automation, Copilot Studio, Power Platform, MCP, Microsoft Fabric, RAGAS, retrieval evals, citation validation, Langfuse tracing, logging, monitoring views, UAT support, technical documentation\nBackend, Data, and APIs: Python, SQL, FastAPI, REST APIs, WebSockets, PostgreSQL, BigQuery, SQLAlchemy, Pydantic, pandas, ETL pipelines, data-quality checks, data modeling\nCloud, Delivery, and Testing: Cloud Run, Cloud Functions, Firestore, Docker, Docker Compose, GitHub Actions, pytest, Playwright, Streamlit, Terraform\nAnalytics and ML Foundation: Power BI, Tableau, scikit-learn, TensorFlow, R, clustering, classification, reliability analytics, KPI analytics, financial analytics\n","permalink":"https://avisheksaha.com/about/","summary":"About Avishek Saha - Applied AI Engineer","title":"About"},{"content":"Currently Reading Coming soon — books and resources that have shaped my thinking on AI engineering.\nRecommended Reading AI \u0026amp; Machine Learning Designing Machine Learning Systems by Chip Huyen — The best practical guide to building production ML systems AI Engineering by Chip Huyen — Essential reading for the modern AI engineer Software Engineering Designing Data-Intensive Applications by Martin Kleppmann — Deep dive into distributed systems and data architecture Online Resources Applied LLMs — Practical guide to building with LLMs MLOps Community — Community resources for ML operations Papers With Code — ML research papers with implementation code ","permalink":"https://avisheksaha.com/books/","summary":"Books I\u0026rsquo;m reading and AI/ML resources I recommend","title":"Books \u0026 Resources"}]