~/lab — interactive systems, no APIs, no recordings
How my systems actually work
No slides, no lectures. The pipeline I run in production, live in your tab — touch everything.
STAGE 01 · INGEST
Semantic chunking playground
why: retrieval quality is decided here, before any model runs3 chunks → 3 vectors
Retrieval starts long before a query arrives. Documents are parsed, cleaned and split into chunks that respect meaning.
…hunks that respect meaning. Each chunk becomes a vector — a point in high-dimensional space. Good chunks keep one idea together. Bad chunks cut a table in half and the answer bleeds out.
… and the answer bleeds out. Overlap stitches context across boundaries so a fact on the edge is never orphaned. At Jagriq this exact tradeoff ran across 36K documents in 22 languages.
Amber = overlap stitching context across edges. Too small starves, too big blurs — this one dial shaped a 36K-document corpus.
STAGE 02 · EMBED + RETRIEVE
Vector space, visible
why: "similar meaning" becomes "nearby point" — that's the whole trickhover to read · click anywhere to drop your own query
My skill graph, projected to 2-D. The ring is the search radius; the lines are top-k. In production: Qdrant, ~329K vectors, HNSW + filters.
STAGE 03 · DECODE
How a model picks its next word
why: every "creative" or "robotic" answer is just this mathcontext: "the drone locked onto the ▁"
// hit sample — 20 real draws from this exact distribution
Real softmax over real logits — struck-through tokens are what top-k / top-p just deleted. Cold = the same word forever; hot = the drone locks onto a sandwich. Every model you've ever used sits somewhere on these three sliders.
STAGE 04 · GROUND
Grounded generation, with the safety off-switch
why: an answer you can't cite is a liability, not an answer[chunk_014] Jagriq pipeline: 36K+ documents, semantic chunking, Qdrant indexing across ~329K structured units and 550K+ entity relationships.
[chunk_087] Citation gate: every source reference is resolved and verified against an external corpus API before an answer is returned.
[chunk_203] Ingestion rebuild: OCR migrated to PyTesseract with async parsing; a race condition corrupting ~22% of documents was found and fixed.
// retrieved chunks on the left become the only things the model may claim
Flip the verifier off and watch it lie in red. At Jagriq this gate is hard: no verified source, no claim.
STAGE 05 · ORCHESTRATE
Multi-agent run, traced live
why: one model is a tool; a coordinated team is a systembus idle
// MCP-style message bus — the pattern behind Memoir's voice automation
Watch the verifier reject and force a retry — that loop is the whole point. Same choreography as Memoir and LearnX, page-sized.
STAGE 06 · LEARN
Pursuit policy, live
why: before LLMs, I taught drones to chase things — act, observe, adjustMEAN REWARD (last 200 steps)
0.00 / 1.0
Crank the noise, watch reward collapse, bring it back. The real thing: SAC + YOLO + VSLAM on a government drone program.