Harit Shukla

Applied AI Lab

Where I build, test and write up what I learn

Hands-on AI work alongside product leadership. Every project carries an honest maturity label, and planned work is labelled as planned.

  • Live in production
  • Prototype
  • Learning build
  • Planned

01 · Live in production

TrueArq Scout

The voice AI agent inside Compass, TrueArq's first product. It holds assessment conversations with students and parents, and a coach uses the resulting report to guide the family. I lead it as AI Product Lead.

Live in productionUsed by 100+ families

On the Scout page

  • How the voice assessment works
  • How we evaluate AI quality
  • What live usage has taught us, and what I chose not to build

Read the Scout story

02 · On GitHub

Existing projects, with source and docs

ai-product-lab (opens in a new tab)
  1. Prototype · Personal use

    AI News Agent

    Python · IBM Bob

    An automated pipeline that fetches, filters, formats and sends a curated AI news digest on a schedule, designed so each stage can be tested on its own.

    • Fetch, filter, format and send stages kept separate
    • Logging, error handling and recovery for long-running jobs
    • Implementation plan and technical spec written before the code

    Source and docs (opens in a new tab)Read the write-up (opens in a new tab)

  2. Prototype · Experimental

    Strategic Decision OS

    TypeScript · Gemini API

    A multi-agent system for structured reasoning about strategic decisions. Agents run in sequence, from framing the decision to an executive synthesis, so each one has the full context.

    • Human review triggered by low confidence, evidence gaps or regulatory keywords
    • Structured, validated outputs and an audit trail for every recommendation
    • Full PRD and architecture documentation

    Source and PRD (opens in a new tab)

  3. Learning build

    Agentic newsletter workflow

    UiPath Maestro Automate

    My hands-on rebuild of a public UiPath demo: an agent drafts a developer newsletter from recent Hacker News stories, and a person approves it before it is sent.

    • Inline human review with Send and Discard routes
    • API retrieval, scripting, agent and email steps in one flow

    Workflow and screenshots (opens in a new tab)

03 · Planned

Next in the lab

Neither project has been built yet. The scope below is proposed, not a description of a working system.

Planned

PV Intake Intelligence

How much AI autonomy is justified in a safety-critical workflow?

Proposed scope

  • Evidence-grounded extraction from adverse-event reports
  • Deterministic rules and human review at the right points
  • Stateful orchestration, mock write-back and audit records
  • Evaluations against clear baselines
Planned

Evaluation & Release Lab

When is an AI system ready to ship?

Proposed scope

  • A reusable evaluation-to-release workflow for real systems
  • Baselines, regressions and failure severity
  • Quality, latency and cost trade-offs
  • An explicit ship or no-ship decision

Scope will be set once the PV Intake evaluation harness exists.

Want to see something running?

I am happy to walk through any of these projects live, including the parts that did not work.

Get in touch LinkedIn (opens in a new tab)