▶ Watch ↗AI Engineer Europe 20261:41:25
Conference affiliation: Unblocked · 2026
Peter Werry is a founding engineer at Unblocked, where he builds context engines for coding agents that connect source code with the conversations, decisions and engineering experience behind it. His work addresses a specific problem: an agent may understand an implementation without knowing which earlier approaches the team rejected or which architectural commitments a change must preserve.
Based in Vancouver, Peter studied at the University of British Columbia and worked at Buddybuild and Apple before joining Unblocked.
By September 2024, his work involved bringing together code and information from Slack, documentation and issue trackers so developers could understand both an implementation and its history. That work has expanded into supplying context for coding agents and automated review. The repository contains the implementation; the reasons for it may live in a discussion elsewhere.
Peter compares an agent to an experienced engineer repeatedly arriving for their first day: it knows how to program but lacks the team’s conventions, previous failures and architectural history. He calls one resulting failure satisfaction of search, borrowing a concept from radiology. An agent finds a plausible explanation and stops looking before encountering an incident report, rejected implementation or conversation that would change its approach. He favors task-specific organizational context assembled across sources, with supporting material an engineer can inspect.
That context also includes who understands a system. Peter describes a social engineering graph built from contribution histories, pull-request reviews and collaboration patterns to locate relevant expertise. Frequent activity alone does not establish authority. Knowledge from private repositories and channels must retain their access restrictions, while conflicting sources and changing documentation make cached generated answers unsafe to treat as permanent truth.
Peter’s teaching on structured query synthesis gives his retrieval argument a concrete mechanism. A request for the pull requests someone merged last week requires identifying the person, selecting merged changes and applying a date range. A semantically similar pull request from three years ago does not answer it.
In his relational context-engine workshop, Peter uses Unblocked’s document-query-engine project to show how natural-language requests become validated MongoDB aggregation pipelines over GitHub issues and pull requests. The model receives a discovered schema, resolves names to identities such as GitHub logins, and proposes a query. The database retrieves the matching records.
The surrounding software limits what that query can do. An allowlist of operators excludes clauses that could execute JavaScript, and hidden metadata fields are removed and reinjected to enforce tenant isolation. Validation and execution errors feed back into a retry loop; full-text search handles cases where exact matching fails. These choices make identity resolution and query validation explicit engineering responsibilities. Peter’s argument is that similarity search supplies only part of retrieval: engineering questions also depend on dates, states, authors and relationships.
Peter’s September 2026 account of Unblocked’s adaptive inference routing explains another part of the team’s infrastructure. The team moved from round-robin distribution to a fixed provider order, then to automatic selection based on observed cost, speed and health. His account preserves the complications that emerged as the system learned from its own traffic:
▶ Watch ↗AI Engineer Europe 20261:41:25
▶ Watch ↗AI Engineer World's Fair 202638:18
▶ Watch ↗AI Engineer World's Fair 202618:35
▶ Watch ↗AI Engineer World's Fair 202619:44
Loading more talks and highlights…