Skip to content
Manoj Gavi

About

I like building AI that's connected to real data, tools, and workflows — not just chat interfaces.

Manoj Gavi

Software Engineer

I'm a software engineer who builds full-stack applications where AI does useful, checkable work. My recent projects take messy inputs — healthcare messages, personal documents — turn them into clean, structured data, and put an AI agent on top that answers questions with evidence and asks before it changes anything.

A chat box on its own is easy to build. The hard and interesting part is everything around it: getting the data into a shape the AI can trust, giving the agent specific tools instead of letting it guess, checking its answers against evidence, and deciding where a human should stay in control.

The engineering story

One approach, two domains

Chapter 1

Healthcare Intelligence Platform

Approaching a complex domain

Healthcare data is a hard place to start: three different formats, strict coding systems, and messages that are often incomplete. Pointing an AI at raw messages would produce confident nonsense.

So I built the boring, important part first — a staged pipeline that parses, validates, normalises, and scores every message, and quarantines anything broken. Only then did I add the agent, with typed tools, a safety screen, and a grounding check that rejects any sentence the evidence doesn't support. An evaluation set keeps it honest as it changes.

Explore the project

Chapter 2

PersonalOps AI

Applying the same architecture to everyday life

The same pattern works for a very different problem: the receipts, bills, warranties, and subscriptions everyone loses track of.

Documents are read and turned into structured records first. The agent then works over those records with specific tools, answers with evidence and a confidence level, and refuses when there's nothing to go on. The new piece is action: the AI can propose a task or reminder, but nothing changes until a person approves it.

Explore the project

How I build

AI should solve a real problem.

It shouldn't be added just because a project uses AI. This is the order I work in.

  1. 01

    Understand the problem

    Who is it for, and what are they doing by hand today?

  2. 02

    Design the workflow

    Map the steps before writing code.

  3. 03

    Build a reliable data layer

    Validate and structure the data first. AI is only as good as what it reads.

  4. 04

    Add AI where useful

    Use AI for the parts rules can't handle — and nowhere else.

  5. 05

    Evaluate

    Test questions with known answers, scored every time.

  6. 06

    Deploy

    Containerised, configurable, cheap to run.

  7. 07

    Improve

    Traces and metrics show what to fix next.

Technical capabilities

What I work with

Only tools used in these projects. Hover or focus any item for a plain-English meaning.

AI & agents

Making AI follow steps, use tools, and show its sources.

  • LangGraph — Runs the AI agent as a clear, step-by-step workflow. Used in: Healthcare, PersonalOps.
  • RAG — The AI looks up relevant documents first, then answers from them. Used in: Healthcare, PersonalOps.
  • Tool calling — The agent uses small, well-defined functions to fetch data. Used in: Healthcare, PersonalOps.
  • LLM APIs — Optional language models (Anthropic, OpenAI, AWS Bedrock) for wording answers. Used in: Healthcare, PersonalOps.
  • MCP — A standard way for other AI tools to use the platform's functions. Used in: Healthcare.
  • Evaluation — A test set of questions that scores answers for accuracy and grounding. Used in: Healthcare, PersonalOps.

Backend & data

Reliable APIs and a clean data layer underneath the AI.

  • Python — Main language for the backends. Used in: Healthcare, PersonalOps.
  • FastAPI — Framework for the web APIs. Used in: Healthcare, PersonalOps.
  • PostgreSQL — The main database. Used in: Healthcare, PersonalOps.
  • pgvector — Lets the database search by meaning, not only exact words. Used in: Healthcare, PersonalOps.
  • SQLAlchemy — Connects Python code to the database safely. Used in: Healthcare, PersonalOps.
  • REST APIs — Standard web endpoints the frontends talk to. Used in: Healthcare, PersonalOps.

Frontend

Interfaces that make complex systems easy to explore.

  • React — Library for building the user interfaces. Used in: Healthcare, PersonalOps, This site.
  • TypeScript — JavaScript with types, which catches mistakes early. Used in: Healthcare, PersonalOps, This site.
  • Next.js — React framework. Used in: PersonalOps, This site.
  • Tailwind CSS — Styling system. Used in: Healthcare, PersonalOps, This site.

Healthcare data

The formats hospitals and clinics actually exchange.

  • HL7 v2 — The classic message format hospital systems send each other. Used in: Healthcare.
  • C-CDA — A standard format for clinical documents like visit summaries. Used in: Healthcare.
  • FHIR R4 — The modern standard for storing and sharing health records. Used in: Healthcare.

Cloud & delivery

Packaging and shipping software cheaply and repeatably.

  • Docker — Packages each service so it runs the same everywhere. Used in: Healthcare, PersonalOps.
  • AWS — Optional Bedrock models in both apps; Route 53 DNS for this site. Used in: Healthcare, PersonalOps, This site.
  • Vercel & Render — Hosting for this site, the app frontends, and the app APIs. Used in: Healthcare, PersonalOps, This site.
  • Terraform — Describes infrastructure — here, the site's DNS — as code. Used in: This site.
  • GitHub Actions — Checks every change to this site before it ships. Used in: This site.

“AI should solve a real problem. It shouldn't be added just because a project uses AI.”

Resume

Want the full picture?

Resume coming soon