What it does

• An idea vault with an API, a web client, and a mobile client sharing one worker system for research, planning, critique, and refinement.
• Agents never overwrite. They propose attributed section changes with version history; the owner accepts or rejects.
• An opt-in discovery mode investigates an idea, records evidence for and against, and applies explicit acceptance rules.

Decisions I’d defend

• PostgreSQL-backed work records with leases, heartbeats, retries, and idempotent scheduling so a multi-hour job can crash and resume.
• Provider routing and usage accounting live outside the workflow, so swapping a model does not touch orchestration.
• Evidence checks and a publication quarantine: a model saying an idea is good is not treated as proof that it is.

Where it stands

Private and running on my machine. Unit and database integration tests cover proposals, workers, migrations, and recovery. Managed staging and mobile device acceptance are still ahead of a public release.

What it does

• A planner splits the question into bounded workstreams, searches iteratively, and extracts full pages and PDFs.
• A claim ledger stores each claim with its source, quoted passage, document location, and a provenance hash.
• Checkpoints in SQLite, cancel and restart, live progress events, a REST API, a dashboard, and a CLI.

Decisions I’d defend

• Two gates, not one: the quote must exist verbatim in the retrieved text, and the claim must be supported by that quote in context.
• Passages are picked for relevance to the sub-question instead of truncating documents from the top.
• When an evidence gate fails the pipeline stops and keeps the rejected material for inspection rather than papering over it.

Where it stands

Private source, working locally. Tests exercise evidence persistence, fabricated-quote rejection, passage selection, and PDF extraction with controlled fixtures. They prove the pipeline behaves; they do not make the internet truthful, so live output still gets read.

Ingestion

• Docling extracts document structure; RapidOCR covers scanned pages.
• A vision model transcribes figures into searchable text, with extracted captions as a fallback when quota runs out.
• Chunks are embedded with a small sentence-transformer into Qdrant, linked back to page and image metadata.

Retrieval

• A LangGraph agent routes between direct answers and retrieval, grades every retrieved chunk for relevance, and rewrites failed queries with a bounded retry.
• Visual questions automatically pull in figure transcriptions; Groq serves the final generation for low latency.
• A no-context refusal path keeps it from inventing answers the document does not contain.

Where it stands

A single-document prototype with in-memory retrieval, built for an assessment. It shows the ingestion and orchestration; persistence and multi-user isolation are out of scope for this version.

The setup

The candidate codes in Monaco. A Python voice agent runs the interview over WebRTC with LiveKit for transport, Deepgram for speech in and out, Silero for voice activity, and Llama 3.3 70B on Groq for reasoning. Next.js in front, Express and MongoDB holding sessions.

Behavior rules

• Socratic: hint, never solve.
• Speaks only on a bug, an approach change, or a stall.
• Cross-checks what the candidate says against what they typed.

After the call

The same agent does a second pass: strict Big-O on time and space, line-by-line fixes, communication analysis quoting the transcript, and a hallucination check comparing claims to code. It renders as a radar chart on the results page.

Falsifiable, not aspirational

• tdd: no production code without a failing test first, and you watch it fail.
• debug: after three failed fixes, question the architecture instead of trying a fourth.
• verify: a claim-to-proof table before “done” is allowed. Each skill names the excuse it forbids.

The memory loop

Lifecycle hooks log every edit and command as compact JSONL. A throttled background distiller extracts trigger-to-action lessons through a strict quality gate, and the highest-confidence ones are injected at the next session start. That subsystem later became its own plugin, engram.

The code graph

A dependency graph of the repo with god nodes, modules, blast radius, and cycles, patched incrementally after each change so it never goes stale.

What it does

• Email-code sign-in, profile editing, saved opportunities, application tracking, reviewable drafts, sharing, export, and account deletion.
• Native subscriptions and credits with StoreKit, server-side purchase records, and recovery of unfinished transactions.
• A clearly labelled demo mode with fictional data so the app can be explored without an account or payment.

Decisions I’d defend

• The mobile cloud service is separate from the desktop sidecar and scopes everything through authenticated endpoints.
• Session credentials sit in the Keychain; account content stays in memory over an ephemeral networking session.
• A purchase is not complete until the server says so.

Where it stands

Compiles and passes demo UI tests on a hosted simulator, including onboarding, tracking, dark mode, and larger text. Real auth, payments, device testing, signing, and TestFlight are still ahead.

Inside the project

• Connects a FastAPI call API to LiveKit voice sessions and a React monitoring interface.
• Explores interruption handling, live transcription, and conversation-event tags. The demo does not establish measured latency or regulatory compliance.

The idea

Every message is a node. Forking from any message spawns a branch that inherits history up to that point, so you can chase a tangent without polluting the main thread. Merging is explicit, and the branch stays intact afterwards so you can keep exploring it.

Smart merging

• Before a merge, Llama 3.3 on Groq summarises the branch, so the parent absorbs only the new insight, not the whole transcript.
• Merge logic filters out shared history and appends only what is new.
• Each merge is recorded as a distinct system event in the chat stream.

Tree navigation

A visual Git-style sidebar renders the whole conversation graph. Click any node to jump; each branch keeps its own history up to the fork point, so switching context is unambiguous.

How it’s built

Next.js App Router with Server Actions, the conversation tree in Neon Postgres through Drizzle, inference wired through the Vercel AI SDK, and full Markdown and LaTeX rendering in the chat.

Inside the project

• Imports public GitHub repositories or locally selected folders into a navigable file graph.
• Uses D3 for node layout and Gemini for contextual file summaries and questions.

Inside the project

• Offers separate hint, tutor, and solution-planning modes in a browser extension.
• Renders Mermaid explanations in a movable, resizable interface alongside the coding problem.

Pipeline

• Vision model (Qwen2-VL-2B via Transformers/PyTorch) ingests each image → structured understanding
• Classifier routes to one of nine categories — anime, tech, hiring, finance, learning, health, news, memes, misc
• Link extractor pulls every URL in the image
• Text cleaner strips hashtags, emojis, formatting noise
• Per-image .txt log in `./output/extracted/` + master `links.md` index

Inside the project

• Attaches wallpaper behind Windows desktop icons and offers a floating scene picker, tray controls, and playback settings.
• Applies WebGL drift, refraction, and lighting to still images, with character-protection masks on selected scenes. These are animated stills, not reconstructed character animation.

Modes

• ASCII text — `<pre>` with dynamically-sized font
• Emoji — `<canvas>` grid with per-glyph bitmap caching for alignment
• Palettes: Basic, Extended, Blocks, Emoji, Custom

Output

Full-color mode samples the original image's pixel colors and applies them per-character, creating the illusion of the original image through typography. Download triggers a dedicated HD render pass at the source image's original resolution — so a 4K input gives you a 4K ASCII PNG, not a screenshot.

Everything runs in the browser

• isomorphic-git on a memfs virtual filesystem
• JSZip packages .git/ + files into a single zip
• No server writes, no GitHub API calls, no fake identities

Inside the project

• Records habit completion in local JSON and derives a daily streak.
• A PyQt mascot changes its visual state with progress.

Color coding

• Blue — BFS explored cells
• Red — DFS explored cells
• Green — Dijkstra explored cells
• Yellow — A* explored cells
• Orange — final shortest path

Inside the project

• Uses a Pygame update loop for mouse movement, collision detection, growth, and scoring.
• A countdown and completion state turn the mechanics into a small playable arcade loop.