// AGENT MODE — STRUCTURED DATA VIEW // or press V to return
Slide 01 — Get data today
Bring your inputs. Tell the agent. Get data today.
Months of assays, finished before the day ends.
Rivulet
Asmay Gharia, PhD · Founder & CEO
asmay@rivulet.bio · rivulet.bio
01 / 08
›
Slide 02 — The data gap · Indication-specific data is the bottleneck
Rivulet
02 / 08
The public data is already exhausted. The future of biology is developing proprietary models fed by your own data-sets.
What the model can already do
10⁹+
virtual compounds enumerated in modern libraries like Enamine REAL and ZINC22
10⁸+
single-cell observations consumed by today's frontier biology foundation models (Tahoe-100M, Arc Institute's State)
Hours
wall-clock to score and rank a full library in silico
What the wet-lab data layer behind it actually delivers
~5,000
cells averaged into a single 384-well readout, with the per-cell distribution discarded
~5%
share of the 7,000-plus known rare diseases with an approved therapy
~0
public single-cell datasets that match a typical rare-disease program's exact cell line and perturbation combination
8–12 wk
wall-clock for a single combinatorial sweep on a customer's specific indication
The model can rank 36 billion molecules. The cell biology it learns from is still averaged across a well. For a rare or indication-specific program, that gap is the program.
‹
›
Slide 03 — Solution · pharma-scale screening on your indication, in a day
Rivulet
03 / 08
A full indication-specific training data set in a day.
The first chip architecture that generates per-cell training data on demand. Any disease, any cell type, any perturbation library. The same flat data shape your foundation model already consumes.
01
ROUTE
Dielectrophoresis steers every particle along the path your experiment plan calls for.
02
READ
The same electrodes capture a live per-particle dielectric signature, along with continuous in-line imagery.
03
EMIT
A per-particle row streams out as the particle exits the chip.
OUTPUT
The flat per-cell table Tahoe, Arc, and Recursion already train on. Anndata, parquet, or whatever shape your downstream pipeline expects. No re-aggregation. No batch waits. No new data formats.
10⁶
particles per run
5 pL
minimum control volume
<1 ms
per-event ML inference
API
MCP + CLI + HTTP agent-callable today
‹
›
Slide 04 — Demo · Your agent calls the hardware
Rivulet
04 / 08
DEMO
Your agent calls the hardware.
01Your agent receives the prompt
02rivulet-mcp returns an ExperimentPlan
03Chip routes, mixes, reads live
04Per-cell rows stream back in real time
05Your agent receives model-ready data
Same agent. Same prompt-to-data loop. Compressed from weeks to hours.
Slide 06 — Use Cases · Six verticals, one validated in an NCI program
Rivulet
06 / 08
Where it is changing the cycle time of a discovery program.
Combinatorial and synergy screens
Drug-drug, drug-target, polypharmacology. The chip co-routes pairs of perturbations with a target cell, reads the per-cell response live, and emits a synergy table by exit.
Compressed from 8–12 wk to a single day
Target ID and validation
CRISPR libraries or compound panels against a phenotypic readout. The chip handles the perturbation-cell pairing, the readout, and the per-cell row.
Compressed from 6–10 wk to a single day
Lead optimization and SAR
Iterative compound rounds with rich phenotypic data per round. The agent feeds the chip the next round, the chip returns single cell data within hours, informing your agent to plan the next experiment.
Compressed from 2–4 wk to a single day
Cell therapy screening (proven in NCI pilot)
Genetically engineered cells screened for editing efficiency, viability, and target expression in a single run.
Compressed from 4–8 wk to a single day
Antibody discovery and B-cell screening
Identify and sort live B cells producing target-binding antibodies, with single-cell impedance signatures separating productive clones from background.
Compressed from 4–6 wk to a single day
Your bespoke wet-lab vertical
The chip is a measurement layer, not an assay-specific instrument, so any wet-lab workflow that needs single-cell, live, programmable readout fits the platform. Bring the assay and we will configure an experiment plan with you.
Scoped in 1 week
Same chip. Same software. Different prompt. Different indication.
‹
›
Slide 07 — Engagement · Pilot Q4 2026, production by end of 2027
Rivulet
07 / 08
Pilot hardware ready to deploy by end of 2026. Production data generation for model training by end of 2027.
Single monthly contract. Cartridges, reagents, platform, and support included. No per-instrument capex. Scale as your run volume grows.
Pilot
Ready to deploy Q4 2026
$50K / month · 3-month minimum
Cartridges, reagents, and support included
Per-cell data delivered into your existing pipeline format
Pilot success criteria agreed up front, written into the SOW
Production
Real data gen for model training by end of 2027
Scoped per deployment
On-premise Rivulet deployment in your lab
MCP, CLI, HTTP API with your authentication
SLA with named technical point of contact
Embedded
2028 onward
Platform fee plus per-interaction usage
Token-style charges per particle interaction processed
Long-tail revenue grows with your run volume
Multi-year platform agreement once the agent loop is your primary integration
Slide 08 — Get Started · Run a pilot by year-end, hardware Q4 2026
Rivulet
08 / 08
Run a pilot on your assays by year-end. Hardware deploying Q4 2026.
Fastest path to evaluating Rivulet is to run one of your real campaigns on the chip alongside your standard process. We provide the platform, the cartridges, and a named technical point of contact. Your team brings the cells, the perturbations, and owns all generated IP.
Path to a pilot
Discovery call: 30 minutes, scoping conversation
Pilot scope written and SOW signed within 2 weeks
Hardware and reagents shipped to customer in order of pilot signing
First per-cell data in your hands within 4 weeks of kickoff
What pilot success looks like
Side-by-side comparison against your current readout on a real campaign
Per-cell data table delivered into your existing pipeline format
Go / no-go recommendation written by your team, not by us
Production data generation for model training: end of 2027.
asmay@rivulet.bio
‹
›
Appendix Gate
12 SLIDES · 5 SECTIONS
Deep dive materials
How it works · the data · use cases · how to deploy · who built it.
A01 — How It Works · Three published primitives on a single chip
Rivulet
A01 / 12
HOW IT WORKS
Three published primitives, integrated on a single chip.
Min volume: 5 pL · Electrode: PEDOT:PSS · Throughput: 106 per run · MCP + CLI + HTTP API
01PEDOT:PSS Electrodes
Conducting-polymer electrode layer enabling both DEP routing and impedance sensing on a single interface. The material that makes simultaneous manipulation and reading possible.
Gharia et al., Science Advances, 2024
02Dielectrophoresis
Electric-field gradients move individual cells along deterministic waypoints. No pumps, no tubes, fully programmable.
Tian et al., Microsystems & Nanoengineering, 2024
03ML Impedance Classification
Multi-frequency electrical fingerprint, cell type and cell state inferred label-free, mid-flight.
Per-cell precision crosses the threshold bulk hardware cannot.
This data is from an early alpha focused on one modality: myeloid cell therapy. In it, we derisked high-throughput single-cell precision and effectiveness above the clinical viability standard. The current device requires no calibration for the assay or task. Same chip architecture. Any biology.
Study: Second-generation myeloid cell therapy for brain tumors and metastatic solid tumors.
National Cancer Institute's Center for Cancer Research · 2025 · Rivulet: pre-clinical validation phase
Condition
Edited
Control
4.74%
Lonza Nucleofector 4D
5.25%
Rivulet
49.1%
Clinical viability threshold: ~20%
Ctrl
Lonza
Rivulet
This is not an incremental improvement. At 5.25%, the best commercial alternative barely outpaces the control and falls far short of clinical viability. At 49%, Rivulet crosses the threshold by 2.5x. The difference is per-cell precision.
‹
›
A03 — The Output · The data shape your models already consume
Rivulet
A03 / 12
THE OUTPUT
The data shape your models already consume.
Plate averages throw away the signal. Per-cell rows preserve it.
Arc's State, Tahoe's atlas, Recursion's maps all consume the same shape: flat, per-cell, per-perturbation rows with provenance. Rivulet preserves that shape by construction. No aggregation, no gating, no means.
· One row per single-cell interaction
· Full provenance: chip → run → plan → waveform
· Drops into anndata, parquet, or the agent's native object format
CLI + HTTP API + MCP server.
run_4812.parquet
particle_id: 0x7fa2
class: "t_cell"
confidence: 0.94
counterparty: ["drug_A", "drug_B"]
reaction_zone: "rz_03"
outcome: "synergy"
duration_ms: 14.2
sort_lane: "lane_0"
plan_id: "plan_4812"
‹
›
Deployment
A04 — Pilot Scope · Pricing and scoping template in one place
Rivulet
A04 / 12
PILOT SCOPE
A pilot scope you can fill in before our second call.
One contract. Cartridges and reagents included. Twelve-week clock.
Pilot pricing (Tier 1)
Line item
Cost
Platform access fee
$50K / month
Cartridges + reagents
Included
Technical support
Included
Custom protocol scoping
Included
Run-volume overage
Quoted
1. The campaign
Indication / target / mechanism: customer fills in
Primary: e.g., distinguish synergy from additivity at single-cell resolution at >X% precision
Secondary: e.g., per-cell data delivered into parquet in <Y latency
Comparison baseline: e.g., current 384-well plate readout
3. Timeline
Kickoff: week 0
First per-cell data: week 4
Mid-pilot review: week 6
Decision point on production: week 10
Pilot close: week 12
4. Deliverables
· Per-cell parquet table for every run, delivered into your S3 or equivalent
· Comparison report (Rivulet vs. baseline) at week 6 and week 12
· Joint go/no-go recommendation by your evaluation lead
‹
›
A05 — Integration · One API surface, nothing else changes
Rivulet
A05 / 12
INTEGRATION
What changes in your stack: nothing. What gets added: one API surface.
Customer system
Integration mode
Status
Agent (LangChain, AutoGen, custom)
MCP server
Available today
Custom orchestration scripts
HTTP API
Available today
CLI workflows
rivulet CLI
Available today
LIMS (Benchling, LabWare, custom)
Webhook + REST
Per-customer
ELN (LabArchives, eLabFTW, custom)
Read-only export
Per-customer
Sequencing handoff (10x, Element, Illumina)
Per-cell barcoded export
Roadmap
‹
›
Use Cases
A06 — Combinatorial · Checkerboard sweep, finished before the day ends
Rivulet
A06 / 12
COMBINATORIAL
Combinatorial screens: a checkerboard sweep, finished before the day ends.
A standard two-compound synergy sweep on 384-well plates runs as a checkerboard of dose pairs. Time per replicate: days. Readout: bulk endpoint per well. Most synergy signal averages out at the well level.
On Rivulet, the same sweep is described as a single experiment plan. The chip co-routes pairs of compounds with target cells, reads per-cell response live, and emits a per-particle synergy table by exit. Time: hours. Readout: per-cell, with the dose pair attached to every particle row.
Standard
8–12 wk
per program
Rivulet
1 day
per sweep
Output schema
compound_pair: "drug_A + drug_B"
cell_state: "apoptotic"
dose_a: 1.0
dose_b: 0.5
synergy_score: 0.91
‹
›
A07 — Target ID · Perturbation libraries on phenotypic readout, in one chip pass
Rivulet
A07 / 12
TARGET ID
Target ID: perturbation libraries on phenotypic readout, in one chip pass.
A CRISPR screen or compound-library perturbation campaign normally requires running the perturbation, waiting on the cell response, dissociating, and then running the readout. Each step is its own cycle, each cycle is its own queue. Total wall-clock: 6 to 10 weeks for a campaign of useful size.
On Rivulet, the perturbation, the cell response, and the per-cell readout collapse into a single chip pass. The library and the cells go in. The per-cell phenotypic table comes out. Downstream sequencing remains an option but is no longer the rate-limit.
Standard
6–10 wk
per campaign
Rivulet
1 day
chip-side readout
Perturbation × cell-state matrix
‹
›
A08 — Lead Optimization · Every round closes by end of day
Rivulet
A08 / 12
LEAD OPTIMIZATION
Lead optimization: every round closes by end of day.
A standard SAR cycle runs round-by-round: synthesize the next set of analogs, run the wet-lab assay, return the data, design the next round. Each round takes 2 to 4 weeks because of the assay. Across 6 to 10 rounds, the wall-clock dominates the project.
On Rivulet, the assay step shrinks to hours. The agent designs the next round at the end of the day, the synthesis queue takes whatever it takes, and the next assay round runs the day the analogs land. The wall-clock for the assay piece of the cycle disappears.
Standard SAR round
2–4 wk
per round
Rivulet assay
1 day
synthesis unchanged
Agent loop with SAR feedback
‹
›
A09 — Antibody Discovery · Live B-cell sorting at chip throughput
Rivulet
A09 / 12
ANTIBODY DISCOVERY
Antibody discovery: live B-cell sorting at chip throughput.
Antibody discovery campaigns currently run through optofluidic platforms (Berkeley Lights, Bruker Cellular Analysis Beacon) or droplet-emulsion FACS workflows. Each runs at throughput in the 104 range, with seven-figure instrument capex and antibody-binding readouts that depend on optical labels.
On Rivulet, B cells are co-routed with a target antigen on chip. The dielectric signature of a productively bound B cell separates from background by impedance, label-free. The chip sorts productive clones into a downstream lane that hands off to standard sequencing or expression workflows. Throughput scales toward 106, label-free, with the same chip used for any other vertical.
Forward-looking
A blinded productive-clone discovery campaign is on the customer roadmap, scoping with two interested partners as of Q2 2026. No paid case study at this throughput yet. Included so the buyer can self-evaluate whether their antibody discovery group should be the first.
B-cell impedance separation
‹
›
Risk
A10 — Deployment Risk · The three questions every customer asks
Rivulet
A10 / 12
DEPLOYMENT RISK
These are the most common deployment questions, and how we answer each.
Risk 1
Platform reliability and uptime
Why it asks itself
A new chip platform inside a production discovery pipeline is a single point of failure if the pipeline is built around it. Customers want to know what happens when the chip is down.
How we handle it
Tier 2 contracts include a 99.5% API uptime SLA, named technical support, and a 4-hour critical-response window. Customers that need higher uptime can deploy redundant instruments. A deployment topology document is provided during the pilot.
Risk 2
Data security, residency, and IP
Why it asks itself
Compound libraries, cell lines, and screening targets are some of the most sensitive assets in the customer's organization. They will not deploy a platform that cannot guarantee data isolation.
How we handle it
Production deployments are on-premise. Customer data does not leave their environment. Classifier fine-tuning runs locally on a customer-isolated model. We do not cross-train across customers. Security and compliance posture available under NDA on request.
Risk 3
Lock-in and exit
Why it asks itself
No customer wants to build a workflow around a vendor that cannot be replaced. They want a clean exit if the relationship ends.
How we handle it
Output is in standard formats (parquet, anndata). All API contracts are documented. Customer data is exportable in full at any time. Cartridge supply, SOPs, and run protocols are in the deployment package. The customer's downstream tooling continues to work whether Rivulet is the assay layer or not.
‹
›
A11 — Your Data, Your IP · Ownership, isolation, and no cross-training
Rivulet
YOUR DATA, YOUR IP
Your science stays yours. Your data never leaves your control.
Production runs on-premise. Pilot runs under a data-handling SOW. We never cross-train across customers.
OWNERSHIP
Anything you run through Rivulet is yours. Every per-cell row, every dose response, every classifier output. We claim no rights to customer-generated data, results, or downstream IP.
ISOLATION
Production deployments live on-premise in your environment. Your compounds, your cell lines, your perturbations stay inside your network. Classifier fine-tuning, when contracted, runs locally on a customer-isolated model.
NO CROSS-TRAINING
We do not aggregate, share, or train shared models on customer data. Each deployment is its own boundary. Security and compliance posture available under NDA on request.
Pilot data handling is defined in the SOW. Discuss residency during the discovery call.
‹
›
The Team
A12 — Team & Advisors · Built by the people who invented the hardware
Rivulet
A12 / 12
TEAM & ADVISORS
Built by the people who invented the hardware.
Asmay Gharia, PhD
Founder & CEO
Six years at Cambridge (EE PhD) inventing the hardware that became Rivulet. Exclusively licensed to Rivulet. First-author, Science Advances 2024.
asmay@rivulet.bio
Katherine Masih, MD, PhD
Co-Founder & CSO
Physician-Scientist Resident at Stanford. Cancer genomics and pediatric cell therapy, the domains behind Rivulet's NCI and Cincinnati Children's pilots. MD, U Miami; PhD, Cambridge.
katherine@rivulet.bio
Austin Weinstein, MBA
Co-Founder & COO
Cancer Entrepreneurship Fellow, Dartmouth Tuck. Owns GTM and commercial.
austin@rivulet.bio
Scientific Advisors
Prof. George Malliaras
Prince Philip Professor of Technology, Cambridge. Bioelectronics Laboratory. World expert on PEDOT:PSS. Spun out 7 biotech companies.
Dr. Iain Fraser
Chief, Signaling Systems Section, NIAID/NIH. Leads the active Rivulet NIAID pilot program.
Bryan Poltilove
Former Head of Cell Therapy, Thermo Fisher. Former Partner, BroadOak Ventures. Advising pharma market entry and commercial partnerships.
