Roadmap

Version 0.1.0 (pre-release), 2026-11-07

The first public release: the search pipelines, the evidence platform, the examiner interface and the case workbench, built and tested. Not yet deployed for a customer.

Released

• Search pipelines. Separate nucleotide and protein pipelines, each search one AWS Step Functions execution. Protein queries are embedded with ESM-2, searched by approximate nearest neighbor in Amazon OpenSearch Service, aligned with MMseqs2 and re-ranked by Smith-Waterman, with profile HMM search and an out-of-distribution check beside them. Nucleotide queries are searched with MMseqs2 on both strands with no length cap, and coding sequences also reach the protein index through six-frame translation. Searches run at two depths, interactive and deep, and many queries can be sent at once through an asynchronous batch API.
• Evidence platform. Every search is pinned to an immutable corpus snapshot and described by a search manifest that can be replayed. Each sequence is stored once with every patent occurrence and its dates. An exact and reverse-complement identity lane runs before any AI tier. Complete result sets are kept and every hit can be paged; each hit names its patent or publication number, its SEQ ID NO and the part of the query it matches, and carries an evidence class. Deep searches apply the metadata filters and the own-filing exclusion, and a result states the known limits that applied to its query.
• Scoring and the examiner report. An empirical null distribution, a Patent Significance Score and a bootstrap false discovery rate for protein searches; PQM and Score-over-Length ranking; codon-aligned display for translated hits. Each search writes an examiner report with its alignment statistics and rendered alignments. The report's narrative is drafted by a language model through Amazon Bedrock and labeled as AI-generated advisory text with the model id; the determination beside it is computed by code.
• Classical sequence tools. Pattern search with EMBOSS, region-constrained exact matches, Needleman-Wunsch pairwise comparison, examiner-built HMM profiles, multiple sequence alignment with a consensus, sequence utilities such as translation and reverse complement, and searches of an examiner's own mini-database.
• Examiner interface. New search, My searches and the result page of a search: evidence lanes, identical-sequence grouping and family masking, ordering and limits, profile hits, unverified candidates kept apart from evidence, notes and relevance judgments, ownership and reassignment, and light and dark color schemes. A search that fails records the stage that failed. The interface is named Seekquence.
• Case workbench. Cases: upload an application's claims, specification and WIPO ST.26 sequence listing; review the claim features a language model proposes and code checks; accept a search plan from a rule-based planner; roll every hit up to references with claim coverage and a documented priority order; group the references and read a checked explanation for each; take rule-based follow-up suggestions; and compose, edit, adopt and close on a cited search summary.
• Case workbench. A narrative of the search summary, drafted by a language model when the examiner asks for one. Code checks each sentence on its own and never shows one it refuses, and the narrative enters the record only when the examiner applies it, labeled as drafted by the model and checked by code.
• Case workbench. The decisions examiners made on closed cases, kept as a dataset inside the customer's account. When a plan is generated, a prior closed case that shares an identical sequence is shown as labeled context. A learned ordering of references is built but switched off and kept outside the evidence chain; it has no measured benefit.
• Examiner tools. A sequence editor with saved sequences, and a shortlist for each case with configurable exports in CSV, JSON, XML, FASTA and tagged PDF, all through one redaction stage.
• Data sources and import. A source-adapter framework with adapters for filed ST.26 sequence listings, NCBI GenBank patent sequences and UniProtKB/Swiss-Prot; a choice of databases for each search; and import of GenCore 6, BLAST and CSV result files into a case.
• Sign-in and security. Sign-in through Amazon Cognito federated to the customer's identity provider by OIDC, with groups mapped to roles; JWT authorization on the API; an application audit log; a network and model boundary with no internet ingress to the pipeline; and secrets in AWS Secrets Manager.
• Security and records. Deletion of case session data. A deletion request by the case's owner or a supervisor hides the case at once, a closed case is deleted after a grace period, and each deletion leaves a record that holds no case material. The searches sent from a case are kept as the examination's search record, and every deletion record lists what stays behind until it expires, such as database backups.
• Security and records. Every export route through the redaction stage, including the search diligence PDF, mini-database FASTA and profile HMM files; an export that would leak withheld content is refused, never delivered. FASTA and profile HMM files go only to internal recipients, and arrive inside a marked wrapper once the customer sets a CUI designation.
• Infrastructure as code. The whole stack as Pulumi code for the customer's AWS account: networking, storage, search, model endpoints, alignment, the API and the user interface. The production stack is defined and tested in code and not yet deployed. A demonstration stack in a commercial AWS Region is defined in code.
• Measurement. A benchmark harness for retrieval quality on public corpora, with a recall-regression gate; a fine-tuning registry with benchmark-gated promotion; and measured gates that decided which channels ship. A nucleotide foundation-model channel, structure-space retrieval, a MinHash channel and rank fusion were built and measured, and are kept off the shipped search path. A mapping of the product's AI use to the NIST AI Risk Management Framework.
• Product site. This public site, its user documentation and this roadmap, generated from the source code at each release.

Planned

• Case workbench. Second versions of the coverage, suggestion and explanation rule sets.
• Examiner tools. Choosing a sequence from a case's sequence listing in the sequence editor.
• Data sources. Adapters for a patent office's own sequence databases.
• Data sources. An adapter for a licensed commercial patent-sequence database.
• Data sources. Adapters for ENA, EST and PIR sequence data.
• Integration. An integration contract, and an adapter for a patent office's document viewer.
• Integration. An adapter for a patent office's sequence-listing status service.
• Integration. Sending and tracking requests to a specialist search team.
• Integration. Sending search records to a patent office's case file system.
• Integration. An NCBI mirror and NCBI BLAST inside the customer's boundary.
• Security and records. A security control inventory for the customer's system security plan.
• Security and records. A software bill of materials and a supply-chain risk assessment.
• Security and records. Transparency documents for AI use: model cards and inputs to an AI use-case inventory.
• Security and records. A published accessibility report for the examiner interface.
• Security and records. Closing the open items that stand before a customer's first production deployment.
• Deployment. Deployment into a customer's AWS account.
• Scale. The production layout of the nucleotide search index.
• Scale. Protein-to-nucleotide search through translated frames, measured at corpus scale.
• Scale. A retrieval gate measured on a reference set of prior art cited by patent examiners.
• Scale. A cost model for running the product.
• Product. A published price schedule and release train.
• Product. A document that draws the line between the shared core and customer tailoring, including the labels each customer configures.
• Product. A trial sandbox.