LTB Buddy: how it was built

No nodes match.

LTB Buddy: how it was builtHow LTB Buddy grew from a first version built in a day into a voice-first tenant assistant with a case copilot, and what it is built to be: calm, plain-language and bounded.This interactive story graph needs JavaScript. The full content follows.

LTB Buddy: A free self-help tool that helps Ontario tenants understand their rights under the Residential Tenancies Act and turn a messy housing problem into an organised next step.

The first version, in a day: The first LTB Companion pages were built in a single day, 17 January 2026.

Conversational intake: Within a day of the first version the app could take a tenant's story in conversation, and soon after it could pre-fill forms from the chat.

Voice as the interface: A voice conversation arrived on 18 January, and on 20 January the app moved to voice as its main interface.

Moving chat and speech to OpenAI: Later in January, chat and then text-to-speech were moved to the OpenAI API.

Reliable structured answers: In February, the way the app turns a conversation into structured facts was made more reliable and was tested.

Seven languages: In March the landing page began highlighting multilingual support, and by May a QA runner covered all seven supported languages.

A filing assistant in Chrome: In April the Chrome assistant gained live detection of portal steps and modals.

Conversations that persist: Conversation history and threads persist, so a tenant can come back to where they left off.

Secure realtime voice: April brought low-latency audio capture and playback, and realtime voice sessions issued securely by the server.

The voice relay: In mid-May the OpenAI voice relay integration was completed.

T6 filing support: In July, support for a safe T6 filing assistant was added.

Grounded in legal content: From the first week the assistant retrieved from a legal knowledge base, starting with T2 and T6 content.

A corpus of LTB forms: A form-ingestion workflow in March was followed by a corpus of 115 ingested documents covering many LTB forms in April.

Measuring retrieval quality: In July retrieval gained an evaluation harness with a quality baseline, so changes can be measured.

Hybrid search: In July lexical and vector search were combined.

Closing the jargon gap: In August a lexical fallback closed a retrieval gap on legal jargon, and the match threshold was tuned.

Watching the law change: A legal-change-watch process tracks 2026 Ontario LTB and RTA changes, a 2027 rent-increase guideline and a pre-release N4 form.

A stateful case copilot: In August the product committed to being a stateful copilot: one case record, not the portal or the chat transcript, is the source of truth.

One case record: A structured case record with an authoritative field registry gives each case a single home.

Reconciling what changes: When an earlier answer changes, the case record reconciles the update and the intake picks it up.

Several issues in one story: A tenant's story can hold more than one issue. Threads can be merged, split and reclassified.

Narratives that stay current: When facts change, narratives that depend on them are marked stale and regenerated, and conflicts with the user's own edits are shown for them to resolve.

Uncertain dates: Fuzzy dates and uncertainty are treated as data rather than as errors.

Complete is not ready: Intake completeness and filing readiness are treated as separate questions.

The assistant reads the portal: The Chrome assistant reports the values it observes in the portal and avoids inserting stale narrative text.

A second mode for clinics: In May the product set out a second way of working for legal clinics: reuse the intake engine as an interviewer and a staff review system.

Clinic admin tools: Clinic panels for managing users and settings, and tools for reviewing chats and assigning intakes, were added in May.

A tamper-evident audit trail: Each generated answer is logged to a tamper-evident chain that stores hashes rather than raw text.

Deleting your data: Deleting your data now also deletes persistent chat threads, and the raw query log was dropped.

An accessibility pass: Heading order, missing top-level headings and landmark gaps were closed in an accessibility pass.

Alpha, openly: In May an alpha banner and a feedback channel were added, so testers know what they are using.

A client-facing brand: Favicon, PWA manifest, 404 page, loading state and email templates were completed in May.

A design system: In June the design system was written down: the interface should feel calm, serious, plain-language and action-oriented.

A self-help boundary: LTB Buddy is self-help software, not a law firm or legal representative, and it measures the workflow progress it can observe rather than asserting legal outcomes.