R&D research agent for F&B innovation

The challenge

The innovation team at a major F&B group in Southeast Asia runs new product development and category strategy, and both depend on knowing what is moving in the market before competitors do. The raw material was there: trade publications, social trend feeds, market data, and competitor launch announcements, spread across more than 100 sources. The problem was the shape of it. Analysts spent their weeks reading, clipping, and pasting fragments into decks, and the same trend would surface five times in five formats before anyone could say with confidence that it was real. By the time a signal had been manually assembled into something a decision could rest on, the window to act on it had often narrowed. The team did not need more information. They needed fewer, better documents.

What we built

We built an autonomous research agent that runs the full pipeline from raw source to finished brief, structured as four stages.

Crawl

The agent continuously ingests the full source set, around the clock. Each source type gets a fetcher suited to its shape: feeds for trade publications, trend APIs for social platforms, structured pulls for market data, and monitors for competitor launch pages. New sources can be added without touching the rest of the pipeline.

Extract

Raw pages and posts are noisy, so the agent converts each item into structured entities and claims: which ingredient, format, brand, or behavior is being discussed, and what is actually being asserted about it. This turns unstructured text into records the later stages can reason over, with every claim keeping a pointer back to its source.

Deduplicate

This stage is where trust is won or lost. A genuine trend echoes across many outlets, and a naive pipeline reports that echo as volume, making one press release look like a movement. The agent merges claims across sources, so the team sees one trend with twelve corroborating sources rather than twelve apparently independent signals. Source counts become honest evidence of breadth instead of an artifact of syndication.

Synthesize

Deduplicated, weighted claims are composed into ranked brief drafts: what the trend is, the evidence behind it, which sources reported it, and why it matters for the portfolio. Briefs land in a consistent structure the innovation team can act on directly, instead of a feed they still have to interpret.

Wrapped around the pipeline is a feedback loop. When the team reacts to briefs, marking what proved useful and what did not, those reactions tune the weights assigned to each source. Outlets that consistently surface signals the team acts on gain influence over ranking; noisy ones fade. This matters for the same reason deduplication does: an agent doing research on your behalf is only useful if you can trust its judgment about evidence, and source weighting makes that judgment explicit, inspectable, and steadily better calibrated to this team’s domain.

How it was delivered

We started with discovery alongside the innovation team, mapping their existing research workflow, cataloguing the sources they actually relied on, and agreeing on what a brief had to contain to be decision-ready. The build followed in increments: crawling and extraction first, so the team could inspect raw structured output early, then deduplication and synthesis once the data foundation held up. We then ran a pilot with the innovation team in their real workflow, using their feedback both to refine the brief format and to seed the source weighting loop. Once the briefs were trusted enough to replace the manual process, the agent moved to production as a continuous service.

What shipped

• An autonomous research agent synthesizing trend and product research from 100+ sources
• A continuous crawl, extract, deduplicate, synthesize pipeline producing ranked, structured briefs
• A feedback loop that tunes source weighting from the innovation team’s reactions
• 70% faster time-to-insight for the F&B innovation team

The innovation team now starts from synthesized, evidenced briefs instead of raw feeds, and the agent gets sharper with every brief they react to.