← Back to portfolio
Case Study · AI Agent · Solo Build

Veridica

An AI research agent that produces verified, source-cited business intelligence — and refuses to be fooled by a rigged source.

Role
Sole designer & builder
Type
Multi-model AI research pipeline
Status
Built & in use (v2.1)
Stack
Node.js / TypeScript · Puppeteer
01

The brief I set myself

Replace the slow, expensive middle of a research team: the reading, cross-checking and writing-up. Point it at a market, a competitor set, or a tender, and get back a finished, sourced report — not a pile of links, not a hallucinated summary.

The hard part isn't fetching data. It's trusting it. A research agent that reads the open web is one poisoned page away from confidently reporting a lie. So the real work was making it verifiable and hard to manipulate.

02

How it works

Veridica runs a multi-stage pipeline: it plans the research, pulls from multiple sources, cross-checks claims against each other, structures the findings, and renders a client-ready PDF with the sources attached. Multiple models are used where each is strongest, and every material claim is traceable back to where it came from.

03

Problems worth solving

Trust

Verified, not asserted

Claims are cross-checked across sources before they make the report, and each is traceable to its origin — so a reader can audit the finding, not just take the model's word.

Security

Dual-layer prompt-injection defence

Fetched web content is untrusted input. A layered defence stops a page that says "ignore your instructions" from hijacking the agent or skewing the findings — the difference between a real research tool and a demo.

Throughput

Hours of work, repeatable

What used to take a researcher days — market sizing, competitor breakdowns, pricing intelligence, tender analysis — runs as a repeatable pipeline that produces a consistent, structured deliverable.

Output

Client-ready deliverables

Finished PDF reports with structure, sources and visuals — the format a business actually pays for, generated end-to-end.

04

What it demonstrates

Agentic AI done responsibly — multi-step planning and tool use, with verification and a real security posture rather than a single prompt. Adversarial thinking — treating web content as hostile input and defending against it. End-to-end delivery — from a research question to a sourced PDF, built and shipped by one person.

05

Built with

Node.js / TypeScript Anthropic Claude Multi-model routing Puppeteer (PDF) Prompt-injection defence Source verification