- Entry date
- 1 August 2025
- Category
- Agents
- Access
- 🔓 Public
Instead of a human scanning industry news for tender opportunities, agents do it continuously — and this is where vector search entered the portfolio for good.
- 1st
- use of vector search (pgvector) anywhere in the portfolio
- Python FastAPI + React 18
- backend/frontend split
- Continuous
- autonomous monitoring, not on-demand
Honest evaluation
Shipped and running; pgvector delivers genuine embedding-space matching rather than keyword search, and the system holds the discovery/decision boundary it set for itself.
What would prove or disprove it further
What would prove or disprove it further: a measured precision/recall check — how many surfaced opportunities a human bid team actually pursued versus discarded as irrelevant — would turn "continuous semantic monitoring works" from a working-system claim into a measured one.
The evidence — full reasoning behind the verdict
Verdict: proven. The claim was that agents could continuously scan railway-industry news and semantically match opportunities against a firm's own capabilities, using real vector search rather than keyword matching — and a shipped, running system doing exactly that, on Python FastAPI/React with pgvector and LangChain, is real evidence for it.
The mechanism is the interesting part: pgvector gives an actual embedding-space comparison between "what this tender needs" and "what we do," which is a materially different (and harder to fake) claim than keyword search dressed up as AI. It's also the first place vector search appears anywhere in the portfolio, and the technique was reused directly in later tender-automation work — a sign the approach held up outside its original context, not just once.
The report also holds a boundary carefully: the agents surface candidate opportunities, they don't decide what to bid on. That's the correct scope for a discovery tool, and the system doesn't overreach past it.
Business development in rail — knowing which tenders are coming, which firms are bidding, which opportunities fit a given capability set — has usually meant a person reading trade press by hand. This project asked whether agents could do the reading continuously instead.
What it was
An intelligent agentic system for monitoring global railway-industry developments and identifying bidding and tender opportunities. AI agents autonomously scan and analyse railway-industry news rather than a person tracking it manually and periodically.
What we built
Shipped & measuredA working system on Python FastAPI with a React 18 frontend, PostgreSQL with the pgvector extension, an LLM API, and LangChain to orchestrate the scanning agents. Opportunities are matched semantically against a firm's own stated capabilities — not keyword search, but a genuine embedding-space comparison of "what this tender needs" against "what we do." That made pgvector the right infrastructure call, and it's the first place vector search shows up anywhere in this portfolio; the technique got reused directly in later tender-automation work.
What we learned — including the honest negative
The real shift here wasn't technical, it was where agents got pointed: business development rather than engineering. Most of the agentic work in the portfolio automates building something; this automates noticing something — continuously watching a stream of industry news no person could keep up with unassisted, and flagging what's relevant.
That's the honest edge worth stating plainly: it's a discovery tool, not a decision-maker. Its value is in continuity — a human doing this manually samples the news; the agents don't.
Where it went / status
Shipped and running as a working system with autonomous monitoring agents. The pgvector semantic-matching pattern proved out here became a reusable building block, carried forward directly into the later (and more ambitious) tender-management system built the same year.
What is still open — kept visible
The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.
- It surfaces candidate opportunities — it does not decide what to bid on; a human still qualifies every lead.
- Business-development scope only: this scans for opportunities, it does not evaluate technical or commercial fit beyond the initial match.
Where this connects
Sources
- vocabotics project audit — RailMonitor (agentic tender scanning, pgvector), Aug 2025vocabotics internal project history · as of August 2025