April 2026
Held-out evaluation became deterministic and place-grouped
We shifted away from tiny convenience slices and toward a cleaner held-out comparison so backbone changes can be judged on the same partition.
Publication notes
Writing that stays close to the deployed system.
This page is the bridge between the working paper, the live route, and the product surfaces people actually use. It collects the notes that matter when the system changes.
What belongs here
Paper-facing summaries of the hybrid route, not just product screenshots and feature announcements.
Honest evaluation notes that explain what improved, what regressed, and where the data or model stack still needs work.
Route-level notes that stay tied to the camera workflow, validation logic, and feedback loop used in the product.
Publication streams
Three threads that now move together.
The paper, the evaluation program, and the public-data training work should be legible in one place instead of being scattered across unrelated pages.
Working paper
Pic2Nav as a production research system
A systems narrative around the hybrid route: direct evidence, retrieval, scene reasoning, confidence gates, and feedback-driven learning.
Read the research pageEvaluation notes
Backbone and held-out comparison updates
Method notes around CLIP, StreetCLIP, deterministic place-grouped validation, and the difference between route utility and model-only scores.
View evaluation notesData notes
Public-data collection and training import discipline
How reviewed public data, city-balanced batches, and canonical training imports fit into the larger NaviSense research loop.
Explore datasetsRecent notes
The publication layer should read like a real research notebook.
April 2026
Nigeria-first public-data import path was added
The dataset builder and reviewed-manifest importer separate collection from canonical training, which keeps public data auditable before it touches the model memory.
April 2026
Public-data promotion is now gated
Approved rows now need image evidence, specific labels, city-level context, and a minimum balanced batch before they are allowed into training.
Ongoing
Publication notes stay tied to the product surface
The goal is to publish what the live system is actually doing, not a separate research story detached from the camera workflow and API surface.
Linked surfaces
Publication notes are one layer of the larger system.
Research page
System architecture, live measurements, backbone experiments, and route-level framing.
OpenBlog archive
Stories, product updates, and the broader writing archive that still lives outside the publication notes layer.
OpenAPI access
The developer-facing surface for the same pipeline described in the paper and evaluation notes.
OpenKeep it grounded
The best publication note is still a tested workflow.
Use the camera flow, inspect the route, and then write from the actual system behavior rather than from memory. That is the bar for the publication layer too.