Johnny Rice

The most effective AI is almost invisible. I connect complex data sources, make them queryable, and deliver fast, accurate answers inside the everyday tools people already know.

ClariTrial is a full example of this principle. It pulls from ClinicalTrials.gov, PubMed, ChEMBL, UniProt, WHO ICTRP, and AACT (a Postgres mirror of the US trial registry), then layers on agentic AI with a graph-based orchestration engine that can compose live queries across those sources in a single conversation. A source health monitor tracks each external API with a circuit-breaker pattern (three consecutive failures in an hour marks a source as down; auto-recovers after one hour). The same underlying data powers the search UI, company pipeline views, competitive landscape pages, an Excel add-in that puts trial intelligence into spreadsheet formulas, and a real-time meeting tool where data-aware agents map what people say to the trial and molecule data the system already knows about.

Before this, the same pattern showed up in informatics roles: Python ETL against vendor APIs, Postgres schemas for compound/assay/batch data, R and Shiny dashboards for scientists (same idea: meet researchers in a tool they trust), Airflow for orchestration, AWS for hosting. The difference now is that the front end is a full product, the data sources are public, and the AI layer replaces some of what used to be manual ad-hoc SQL sessions. The engineering habits are the same: normalize carefully, treat APIs as contracts, cache aggressively, and keep the path from raw records to what the user sees explicit enough that you can debug it six months later.

The core product is a clinical trial intelligence tool. You can search 470K+ trials, browse company pipelines, compare trials side by side, and ask an AI chat questions that hit live databases instead of a static knowledge cutoff. The AI orchestration layer uses a lead model that delegates to specialist agents (trial discovery, evidence synthesis, deep dive, comparison, and CDD Vault agents), each scoped to a specific data source with its own tool call budget, step limit, and reflection loops. A dynamic orchestration engine classifies queries as simple, moderate, complex, or debate-worthy, builds a dependency graph, and dispatches agents in topological order with parallel execution where dependencies allow. Every tool result carries provenance metadata (source, timestamp, row counts), and the full trace is visible in the UI: which tools ran, per-tool latency, specialist model IDs, tool budget usage, and whether a specialist was retried after a reflection failure.

Data access is governed by layered guardrails. A prompt guard runs moderation and relevance checks before the model sees user input. AACT queries use either allowlisted SQL presets (four fixed modality slices) or parameterized flexible queries where all user-supplied values are bound as $N parameters. An injection detection layer scans filter values for suspicious SQL patterns before queries execute. Results go through a validation pass for empty sets, implausible counts, and missing fields. Every chat turn is logged to a JSONL audit file with prompt version, model ID, and the serialized tool trace.

The part highly relevant to drug discovery use cases is how many data types connect. A single conversation can pull a trial from ClinicalTrials.gov, cross-reference the sponsor's pipeline from curated company data, check bioactivity on the drug target from ChEMBL, look up the protein structure from UniProt, and query AACT for competing trials in the same indication and phase. The same entities (companies, drugs, targets, conditions) persist in an entity memory layer (a lightweight knowledge graph in Postgres) so the system remembers what you've looked at across sessions. Cross-session context is injected into the system prompt when the user mentions previously seen entities.

Meetings are the clearest example of zero-friction AI. Deepgram Nova-3 handles real-time speech-to-text over a browser WebSocket, with domain-specific keyword boosting (molecule names, gene targets, clinical terms) so the transcript is actually useful for a research discussion, not garbled. Three specialist agents (infrastructure architect, data scientist, cheminformatician) are data-aware and tool-aware: they listen to the conversation, recognize the molecules and targets being mentioned, and recommend relevant tools from a 70+ catalog spanning AWS, cheminformatics, proteomics, and genomics. Nobody stops the discussion to search a database. A batch analysis loop runs every 20 seconds, extracting entities and mapping them to the trial and molecule data ClariTrial already indexes. The post-meeting summary ties everything together: what was discussed, which data points are relevant, and what to follow up on.

There is also a mission flow (multi-agent task runner) where you can describe a research question, the system generates an execution plan, and specialist agents run in parallel via SSE streaming. Results are synthesized into structured reports with confidence scoring aggregated by source tier and recency. For contested questions, a multi-agent debate protocol runs three perspectives (optimistic, skeptical, balanced) through challenge rounds before a judge synthesizes the final assessment. An agent capability registry declares each specialist's capabilities, cost tier, and step budget so the orchestrator can plan execution graphs without hard-coding agent selection.

The CDD Vault integration bridges internal discovery data (molecules, batches, assay protocols) with external clinical intelligence. Credentials are encrypted at rest with AES-256-GCM, permissions are role-based (five tiers from clinical analyst to vault admin), and the system scaffolds custom agent teams from a vault's schema. Scientists stay in their existing vault workflow; the AI reads the schema and adapts to it rather than forcing a new interface. Write operations require two-step approval with HMAC-signed tokens, and every write is audited to both an in-memory buffer and a Postgres table.

The Excel add-in is the same idea applied to the most familiar tool in any organization. Custom formulas like =CLARI.TRIAL() and =CLARI.ASK() put trial lookups, pipeline data, and natural-language AI queries directly into spreadsheet cells. A task pane lets users search trials and insert results without leaving Excel. The integration is nearly invisible: the same data layer that powers the chat, the meeting tool, and the mission runner also backs a set of spreadsheet formulas. End users get powerful access to multi-source clinical intelligence through a tool they already open every morning, with no new workflow to learn.