Ziyun Liao

parse · deterministic-first

• The parser is treated as an untrusted component: every parse is audited against the source PDF's own text layer, then repaired — or refused — before anything downstream sees it
• Every page's provenance is recorded — native text layer vs image-derived — so downstream stages decline to reason over content the parser only guessed at

index · hybrid + pre-computed

• Chunking follows the document's own clause-and-section grammar, so a retrieval hit lands on a citable contractual unit, not an arbitrary span sharing vocabulary
• Vectors are deliberately not the only index: typed facts (amounts, dates, cross-references) and structural anchors are extracted deterministically at index time — removing an entire class of wrong-but-plausible evidence

retrieve · evidence-first grounding

• Retrieval is a recall mechanism, never an arbiter: nearest neighbors always return something, so structural evidence outranks similarity hits — no finding is decided on similarity alone
• Lexical and semantic signals are fused and re-ordered before any LLM sees the evidence — a choice driven by measurement: on real corpora, raw scores and misleading evidence overlapped completely, so no cutoff could separate them

geometry · 360° → 12-pinhole rig

• Equirect panoramas reprojected into 12 rigid pinhole views (4 yaw × 3 pitch) so standard SfM applies — with fixed rig extrinsics massively constraining the solution
• Geometric masks exclude the operator, zenith/nadir and seams from ever polluting features

features · learned over handcrafted

• DISK learned features with LightGlue matching — robust on the low-texture, repetitive surfaces of construction sites where SIFT gives up
• Sequential windows plus loop-closure pairing, geometrically verified before entering the map

solver · GPU bundle adjustment

• Incremental SfM with cuDSS sparse-direct bundle adjustment on GPU (~2.3×) — intrinsics and rig extrinsics held fixed throughout
• Continuation pass recovers unregistered frames instead of silently dropping them

contract · sim2 to plan space

• Two field-anchored points solve scale, rotation and translation from dimensionless SfM space into real floor-plan pixels — with mirror-flip detection
• Every hotpoint ships with heading, in-frame ratio and a confidence gate, not just coordinates

state machine · lesson as backbone

• The teaching loop is modeled as an explicit state machine — scheduled → live → organizing → publish → grading → completed — with tools and permissions gated per phase
• AI suggests, teacher finalizes: every AI artifact (slides, summaries, grades) stays a draft until a human ships it

agent boundary · MCP orchestration

• PPT generation runs as a design-first agent over stdio MCP, rendered asynchronously in the background while the teacher keeps working — then edited page by page with the agent
• Tool calls are constrained by whitelist and phase — the agent can only act inside the current lesson state

delivery · desktop-first, web-ready

• Electron with strict IPC isolation: the renderer never touches keys, paths or processes — model configuration stays in the main process
• TypeScript monorepo on a Supabase-aligned schema, so the persistence swap is a migration, not a rewrite

orchestration · multi-provider LLM routing

• Admin-managed provider catalog with model discovery and per-deployment defaults — providers swap without touching product code
• A constraint-solving timetable engine negotiates coverage across 80 teachers × 24 classes under hard constraints