Open Hub

Development journal

The Evopien AI and robotics development journal.

The journal is the chronological engineering record behind the public project - hardware, software, decisions, failures, and the limit of every result.

Latest update / J-020

Introducing Evopien Hub, a private workspace for your day

Plan your day, organise projects, connect email and calendars and receive reminders in your private Telegram chat. Evopien Hub opens with an invitation-only beta.

Read J-020

Latest first

Real progress, with the known limit kept beside it.

20 records shown for All records.

J-020Web workspace beta
Evopien Hub introduction for capture, tasks and calendar setup

Introducing Evopien Hub, a private workspace for your day

Plan your day, organise projects, connect email and calendars and receive reminders in your private Telegram chat. Evopien Hub opens with an invitation-only beta.

Completed
A newly invited account signed in to Hub, attached Gmail and Outlook, and imported both calendars. Separate live tests exercised summaries and reminders in Telegram.
Known limit
Invitation-only access. Summaries are optional, and calendar imports run on request. Saved connections are deleted in One.
RELATEDExplore Evopien Hub
J-019Physical presence and local interaction · Historical
Studio visualization based on Evopien’s assembled physical head, with its camera and illuminated display

Evopien moves. Remembers. Responds.

A moving head. Local voice and vision. Voice onset measured at a 1.35-second median in a recorded replay, web answers around 1.6 seconds, and governed memory tested across 98 simulated days. Watch the English and Spanish demos and meet the hardware bringing Evopien into the room.

Completed
The 5 November demo brings physical face following, visual watches and spoken counters together. Separately, governed memory has reached 98 simulated days and 2,200 staged interactions, with 129/129 reviewed answers in the latest week.
RELATEDWatch the demos
J-018Governed memory and audio/video research · Historical
AI-generated concept of a glowing memory timeline with sound waves and a visual-context motif

98 simulated days: Evopien’s next leap in memory and conversation

Nearly 100 simulated days. 2,200 synthetic interactions. A new research milestone for memory that follows changing facts, preserves the meaning of preferences and remains under owner control, alongside progress toward natural audio/video conversation.

Completed
The original campaign reached simulated Day 98. Its latest weekly batch passed 129 answer checks plus state checks. Audio/video components have advanced within their measured test boundaries.
Known limit
98 simulated days, not 98 elapsed days or a final-source replay of all 2,200 inputs. Historical failures remain recorded; six-month, year and full natural audiovisual-duplex acceptance remain unfinished.
RELATEDInspect current development
J-017Working demo and governed memory · Historical
AI-generated concept of Evopien with fading conversation and deliberately saved notes

Governed memory in everyday use: a demo worth preserving

Temporary conversation context, explicit saved preferences and useful notes meet English/Spanish speech, camera input and public web research. We preserve an accepted working demo, explain what owner control looks like, and thank Seeed Studio for the hardware helping bring Evopien to life.

Completed
Two software rehearsals exercised saved memory, correction, selective deletion and chat/database restarts. The founder accepted the live video baseline with remaining bugs; microphone, Gemini camera, speaker and display are integrated on the bench.
Known limit
A working demonstration, not a finished product or full Alpha acceptance. Some natural-language edits and speech recognition still need repair; response latency and broader security, recovery and reliability gates remain open. The physical head and neck are not assembled.
RELATEDExplore the current demo
J-016Hardware arrival and collaboration · Historical
Delivered Seeed Studio components, including reSpeaker Flex, Orbbec Gemini 2, servos and a display, laid out before integration

Building the Evopien head with Seeed Studio: the hardware has arrived

The components from our collaboration with Seeed Studio are here, and so is our Bambu Lab A1. A thank-you to the team helping us build the first physical head, a look at the concept designs, and the next steps after the Hindsight memory integration.

Completed
Seeed Studio components received and photographed; Bambu Lab A1 received and set up in the workshop.
Known limit
Hardware arrival is preparation, not a working head. Hindsight integration and testing are in progress. The head and body images are concepts; the real build may differ. Next week’s hardware start is a target, subject to the remaining checks.
RELATEDCurrent development
J-015Governed memory checkpoint · Historical
Original conceptual illustration of separate memory records and permission boundaries around one Core

From a working demo to governed memory: one week building the next layer

The 9–16 September engineering review: local/cloud demo recovery, durable synthetic memory, reviewed English/Spanish input and governed speech. The latest operator-reported backend run passed 903 required methods; broader cognition, voice and Alpha gates remain open.

Completed
Synthetic PostgreSQL persistence, current-rights checks, reviewed correction/deletion and bilingual task/time workflows across restart; bounded physical interruption recovery on its earlier exact source.
Known limit
Fictional-data engineering evidence, not a real-user Alpha. A dated update records the narrow bilingual recovery diagnostic; full G8/G9, deployment security and independent recovery remain open. Actual-model probes retain failures. Engineering main still preserves the separate demo freeze.
RELATEDSee current development
J-014Integrated demo freeze · Historical
AI-generated conceptual artwork of sound, vision and an internet constellation converging around one luminous core

Voice, vision and internet: one frozen foundation, local and cloud

424 engineering commits bring full-duplex conversation, selected live and recent-history vision, and public research into one Thor demo. The baseline is frozen and merged into main, with local and cloud profiles. Governed memory and stronger permission boundaries are next.

Completed
One console and terminal runtime connect bilingual speech, interruption, pause/resume, selected vision, local James/DDG retrieval and optional Groq/Exa services. The September 9 cloud demo exercised these paths and stopped cleanly.
Known limit
Founder-accepted engineering freeze, not formal acceptance. MS-07, MS-08 and MS-09 remain open; routing defects, source freshness, latency tails and complete latest-local-profile qualification remain. Session recall is not governed durable memory.
RELATEDSee the current development state
J-013Thor voice + vision checkpoint · Historical
AI-generated conceptual illustration of camera frames and an acoustic waveform meeting in one core

Bilingual voice meets visual context: the next Thor checkpoint

Explicit English and Spain-Spanish switching is frozen, continuous perception has been exercised, and Qwen has received a current camera image. The latest branch adds a 15-second visual-context window to ordinary turns, with physical qualification still ahead.

Completed
235 commits since the previous engineering cutoff cover conversation continuity, long questions, short follow-ups, explicit bilingual control, C920 recovery and freshness, adaptive perception with PVA tracking, current-image grounding, and bounded recent-video experiments.
Known limit
MS-08 and MS-09 acceptance gates remain open. CP09-S measured 1.524 seconds median speech-end to first generated PCM, with 1.803 seconds P95. Recent-video motion tests failed their rubric; the 15-second candidate has not yet been physically qualified, and integrated voice/video latency is unmeasured.
RELATEDSee the current development state
J-012Thor engineering checkpoint · Historical

Five days rebuilding the Thor voice loop: a 1.37-second median and the gates still open

Since J-011, Evopien has moved from separate realtime components to a physically exercised Thor speech path, frozen a Core-owned conversation lifecycle, re-architected the Qwen history path so the old cache cliff did not recur in the 100-prompt text-only test or short eight-response physical run, and reduced median speech-end-to-first-PCM latency to 1.3705 seconds.

Completed
ConversationLedger now owns spoken-only history, WAIT/CONTINUE/STOP and generation leases; Smart Turn overlaps Cohere transcription; a derived open-session memory index is separated from a bounded cache-stable Qwen working set; and one natural Qwen stream now feeds locally framed Scylla speech through a single playback owner. A 100,000-turn synthetic-history text-only qualification held later Qwen TTFT at 244.3 ms P50, and the latest physical run held Qwen TTFT near 212 ms P50.
Known limit
The 1.3705-second result is a median from detected speech end to first generated PCM, not an independently measured acoustic hearing timestamp. P95 was 1.7971 seconds and failed the ≤1.3-second gate; only eight Qwen turns were measured, false-barge cleanup and temporary generated residue remained open, and the paired bilingual natural-duplex exit corpus has not passed.
RELATEDSee the current development state
J-011Thor engineering checkpoint · Historical
Closed NVIDIA Jetson AGX Thor Developer Kit box on a round table

Thor arrived: what worked, what failed, and the realtime stack now being integrated

Since the pre-arrival record, Jetson AGX Thor has been brought up, the standalone audio and camera foundations have been reproduced, a bounded Qwen3-Omni reference has been completed and retired from active conversation, and the governed Core identity gate has passed. Evopien is now integrating a modular English and Spain-Spanish realtime stack.

Completed
The Thor platform is operational; MS-03 passed for its declared XVF/Pebble AEC and C920 fixtures; MS-04 received a conditional pass for its exact fixed-corpus bilingual multimodal reference; and MS-06 completed one creation-with-lineage event. Nemotron ASR, turn-intelligence, Qwen3.8/SGLang, Magpie/NanoCodec, and cancellable playback have since been exercised as separate experimental slices.
Known limit
MS-08 remains open. The complete ASR-to-Core-to-cognition-to-speech loop is not integrated or validated, physical acoustic stop latency and Castilian voice quality remain open, and MS-09 continuous scene-state vision has not started. This is not Alpha, production, internet, or powered-motion readiness.
RELATEDSee the current development state
J-010Architecture transition · Historical

From Orin Nano to Jetson AGX Thor: rebuilding the backbone before the machine arrives

Evopien is moving from its first Orin Nano prototype to a clean-room Jetson AGX Thor programme. The earlier implementation is pinned as historical engineering evidence; the Thor generation starts from governed requirements, explicit verification gates, and a new identity with disclosed lineage rather than inherited runtime state.

Completed
A new Thor-native project baseline now defines 102 requirements, 16 evidence-gated milestones, 30 tracked risks, source authority, identity lineage, fail-closed interfaces, and a hardware-arrival sequence. The active repository passed its host-independent baseline checks before publication.
Known limit
The Jetson AGX Thor has not yet been received or verified. Qwen3-Omni is only the first local-provider candidate; no Thor model, audio, camera, networking, performance, memory, identity-activation, or physical-behaviour result is claimed. Founder-review documents and release gates remain open.
RELATEDSee the Thor development state
J-009Engineering milestone · Historical

Voice V1 is frozen: the next milestone is conversational presence

The Orin-generation experimental Voice V1 was frozen after representative clean physical turns reached about 2.15 seconds from user end-of-speech to first meaningful audio, with a fastest retained turn near 1.51 seconds. This remains historical Orin evidence, not the current Thor runtime.

Completed
At that checkpoint, the bilingual voice baseline combined persistent six-thread Parakeet ASR, Qwen3.5-4B, selective-INT8 Piper speech, XVF3800 hardware echo cancellation, hardware-AEC-assisted barge-in, qualified endpointing, 233 passing tests plus two subtests, and a tagged recovery anchor.
Known limit
About two seconds is a representative result, not a guaranteed ceiling. Conversational Presence R0 is still research: no vocal backchannel, visual nod, reliable spontaneous end-of-turn policy, robust multi-speaker ownership, governed persistent memory, or physical actuation is claimed yet.
RELATEDSee the current development state
J-008Engineering checkpoint · Historical

From local voice to an integrated head: the 14-17 August build

Between 14 and 17 August, the 8 GB Orin Nano head gained hardware-AEC-assisted simultaneous capture/playback, bilingual interruption, Silero speech qualification, fresh on-demand visual turns, a DeepStream/RF-DETR/NvDCF perception mode, and observe-only direction telemetry. The retained suite reached 188 passing tests.

Completed
That local head could hear, answer, be interrupted, switch between English and Spanish, capture a fresh C920 frame for visual questions, and run a separately preserved continuous detector/tracker pipeline. Direction-of-arrival telemetry was integrated without allowing unreliable direction data to reject speech.
Known limit
Robust multi-speaker ownership is not solved; DoA produced sustained wrong directions in live testing. Governed persistent memory, the complete multi-user privacy proof, fluid low-swap performance, launcher lifecycle hardening, audio-video active-speaker fusion, and physical actuation also remain unfinished.
RELATEDSee the current development state
J-007Working milestone · Historical
Illustrative reconstruction of the current Evopien Jetson voice-development workstation with microphone array, compute board, and speakers

Evopien can now hear, answer and be interrupted

At this Orin-generation milestone, the first Evopien head prototype ran a local voice loop: microphone input, speech recognition, local Qwen reasoning, streamed speech synthesis, hands-free turn-taking, and barge-in cancellation.

Completed
Local microphone capture, whisper.cpp transcription, Qwen3.5-4B reasoning, Piper speech, streaming output, hands-free turn-taking, and barge-in interruption worked together through the archived Orin-generation Core runtime.
Known limit
The C920 camera and Qwen vision work independently, but live conversational vision is not connected yet. Governed persistent memory, the full multi-user privacy proof, ESP32 behaviour, and physical actuation also remain future milestones.
RELATEDSee the current development state
J-006Project record · Historical
The first Evopien Jetson workstation during the working head and Core milestone

The first canonical Evopien identity now persists across repeated runs

The Orin head platform established GPU acceleration, live camera capture, microphone input, speaker output, and a local Evopien Core software foundation. Its predecessor identity was created in persistent storage and returned unchanged across repeated application runs.

Completed
Head-platform I/O and GPU execution were validated, the Core API and persistent database stack ran, and the predecessor identity was retrieved across repeated bootstrap runs instead of being recreated. It is not imported or activated on Thor.
Known limit
This record predates the local voice milestone. Governed persistent memory, the complete multi-user privacy proof, conversational vision, audit/governance memory flows, and physical actuation remain to be implemented and tested.
J-005Project record · Historical
NVIDIA Jetson Orin Nano Super used for the first Evopien head workstation

JetPack 7.2 is running on the first Jetson head workstation

JetPack 7.2 with Jetson Linux 39.2 ran from the 1 TB NVMe drive on the first Jetson Orin Nano Super workstation. This historical record documents the ISO installation, first-boot issue, system state, and then-open validation work.

Completed
The official JetPack ISO was installed to NVMe, Ubuntu 24.04 reached the desktop, the NVIDIA package pass completed, and the first workstation record was created. Exact CUDA, Python, Docker, audio, and camera checks remain explicit follow-up items.
Known limit
This is a setup and implementation record, not yet evidence of an integrated head, Milestone M1, or a validated physical prototype.
J-004Project record · Historical

Repository created and implementation started

Project control moved from planning into an active private repository and initial implementation work began.

Completed
The repository exists and development has started.
Known limit
No public commit reference or M1 test result is attached, so this remains a project record rather than demonstrated evidence.
J-003Project record · Historical

Core prototyping hardware purchased and placed in transit

The Jetson compute platform, ESP32 controller, breadboards, measurement tools, power equipment, and supporting prototyping supplies were purchased for the first stationary-head programme.

Completed
Core hardware was purchased and entered the delivery phase.
Known limit
Purchasing equipment is preparation, not proof of a physical prototype.
J-002Controlling draft · Historical

The first build was narrowed to a credible proof

Alpha v0.1 defined the smallest system worth building and the gates required before partnership readiness.

Completed
Requirements, stages, release gates, and M1 were drafted.
Known limit
Alpha v0.1 remains a founder-review draft and its requirements remain unproven.
How evidence labels work
Decision
An approved direction or controlling principle.
Project record
A dated engineering or setup update.
Prototype
A working implementation in a defined test environment.
Demonstrated
A behaviour observed against a stated test.
Historical
A dated result preserved from a superseded generation; not current-system evidence.