Ryan Tang
BenchSciLead Product Designer2025

Agentic tables and chat differentiation for scientists workflow

Led a prototype exploring how AI could reshape how scientists interact with research data. Discovered an unexpected usage pattern during testing, validated the concept with scientists and commercial partners, and shaped it into a meaningful direction for the platform.

PrototypeAgentic orchestrationCursorClaude CodeInnovation
Shipped prototype hero: table-aware interaction where scientists add columns and inspect evidence in context.
Roadmapped
Differentiated AI Experience
0+
Research sessions
0%
Unprompted adoption
Differentiation
New AI interaction model

Led a prototype to test one interaction pattern: scientists diverging from evidence tables, not from chat.
That single pattern became a roadmapped, differentiated AI direction.

Final interaction: table-led divergence with visible progress, context continuity, and table-aware follow-up actions.

Observation + opportunity

MENTAL MODEL

Where scientists actually diverge

The assumption

Scientists diverge in chat

Follow-up questions happen after reaching a decision

Query

e.g. TLR4 IHC Protocols

Response

AI directs

Tool

Evidence table

Decision

e.g. protocol conditions

Follow Up

New question

"What primary antibodies would work best?"

The observation

Scientists diverge at the table

Follow-up questions happen before deciding — driven by what they see in the evidence

Query

e.g. TLR4 IHC Protocols

Response

AI directs

Tool

Diverge here ↓

Primary antibodies performance

Follow Up

New question at table

Decision

Primary antibody + conditions

"What secondary antibodies would work best?"

Scientists kept trying to branch while still inside the evidence table.
Opportunity: make table divergence first-class so context stays anchored and follow-up work feels continuous.

  • —15+ sessions: validated with scientists and internal partners.
  • —100% unprompted behavior: participants reached for table actions without instruction.
  • —Roadmapped direction: adopted as a differentiated AI interaction model.

Persistent context bus

Agentic table architecture

Mini data pipeline — original intent preserved at every step

Pipeline steps

01User Input
02Intent
03Primary Driver
04Supporting
05Data
06Ranking
↕
↕
↕
↕
↕
↕
Memory (Context Bus)
queryintentprimaryDrivercolumnsrowstableDatasummary

Outputs

→
Same step · parallel
Table Output
Summary
Column & template
Add Column ↺
Templates ↺
processQuery(query)

The scientist's natural-language question enters and is pinned to the bus for every downstream step.

reads: —writes: query

Origin query context persisted across add-column, filter, and chart actions so scientists did not re-anchor every step.

Witnessed reasoning (visible pipeline)

Pipeline state remained visible while new columns and insights were generated.

The system showed where it was in the orchestration path, making intermediate states inspectable instead of opaque.

Streaming with parallel summary

While evidence views updated, summary signals streamed in parallel to reduce dead time.

Structured streaming made latency feel productive: the table kept moving while synthesis caught up.

Under-tuned prompts for input flexibility

Loose natural-language input could still resolve into table-scoped actions.

Prompts were intentionally under-constrained so scientists could phrase requests naturally without breaking table intent.

Template-seeded prefetch

Template actions seeded likely next queries to make divergence faster and more predictable.

Reusable templates warmed the next state and reduced cold-start friction during exploratory loops.

Reflection

Coding the prototype was the leverage point: it exposed real orchestration and latency behavior quickly, so design decisions were based on working system dynamics, not storyboard assumptions.

← All work