An agent that does the digging, over 25 million records.
An infrastructure-intelligence platform with an AI agent that plans a question, calls the right tools in sequence, and assembles the answer.
- Role
- Frontend engineer
- Company
- Taiyo.AI
- Where
- San Francisco (remote)
- Scale
- 25M+ records, 50+ clients
- Year
- 2023 — 2024
Infrastructure intelligence, and the time it took.
Taiyo.AI sells intelligence on infrastructure projects — tenders, contracts, spend — to enterprise clients deciding where to bid. The value is in the dataset: tens of millions of records, across countries and languages.
The problem was never whether the answer was in there. It was how long it took an analyst to assemble it: roughly four hours of searching, filtering and cross-referencing for a single usable insight.
The things that made it hard.
Analysts were not short of data. They were short of the hours it took to assemble it.
One prompt, one answer does not survive a question with several steps in it.
At 25M records, latency is the product — a correct answer after five seconds is still a bad experience.
How do you answer a question that needs five searches, not one?
Real questions decompose. 'Which water projects in these three states slipped last quarter, and who won them' is a search, a filter, a join and a summary. A single prompt against a single index answers none of it.
3 passes at it.
Make the search fast enough to build on
- Profiled the Elasticsearch queries behind the slowest screens and found the bottlenecks were in how the queries were shaped, not the cluster size.
- Designed multi-tier caching so repeated and near-repeated queries do not touch the index.
An agent that calls tools, not a chatbot
- Exposed the platform's own capabilities — search, filter, aggregate, summarise — as tools the model can call.
- The agent plans a question into steps, calls each tool in turn, and assembles the result from what comes back.
- Every step is inspectable, so an analyst can see which searches produced the answer rather than trusting a paragraph.
Making the generated text trustworthy
- Built a Gen-AI news module on OpenAI streaming APIs for 100+ users.
- A/B-tested prompt variants against a fixed question set rather than tuning by feel.
- Cut token costs 45% by trimming context to what each step actually needed.
What came out of it.
The agent was the interesting part. Letting a model call the platform's own tools, in an order it chooses, turned a search product into something that answers the question you actually asked.