L-01
agentic systems · llm evals · production
FriesenLabs
My applied-AI lab. The flagship is Uplift, a multi-tenant agentic CRM: describe what you want, and agents build and run it — your whole business becomes something you can chat with, ask questions of, and hand work to. Hardened the way agent products usually aren't: a 200-question graded eval harness, failures root-caused into a taxonomy, fix waves measured against baseline.
Multi-agent
Bedrock · Aurora · ECS
200-q eval harness
Multi-tenant
RAG "brain"
L-02
multi-agent orchestration · dev infra
agentFleet
The build system behind most of this page. A boss agent decomposes a goal into file-disjoint tasks and fans them out to tiered subagents in isolated git worktrees, spread across several Claude Code and Codex accounts on a small fleet of Macs. Every result faces an adversarial, refute-by-default review panel before a serial squash merge keeps main green. Foreman agents work through multi-brief chains overnight, unattended.
Boss / foreman / builders
Git worktrees
Refute-by-default review
L-03
vector search · HNSW · retrieval systems
HNSW Retrieval Research
Two studies into how approximate-nearest-neighbor indexes (HNSW) actually behave at scale, the engine underneath every RAG and semantic-search system. Predict, then Probe learns to set each query's search budget instead of guessing one globally, holding recall while cutting work; a companion result isolates a failure mode where piles of exact-duplicate vectors quietly collapse an index's reachability, cured by a tiny epsilon-jitter that restores recall at ~33% lower cost. Lab-measured, replicated in the wild, novelty confirmed against the literature, with an interactive calculator.
HNSW · ANN
Recall vs cost
−33% cost
Interactive calculator
L-04
music retrieval · contrastive learning · from scratch
ChordSpace
A similarity embedder for chord progressions — a model category that, per a systematic sweep, didn't exist. Built fully from scratch: custom parser, transposition-invariant encoding, and a 2.3M-param transformer contrastively trained on 679k progressions. On a leakage-controlled retrieval eval it already beats the classical baselines (TF-IDF, edit distance); a hybrid of both wins every eval. In progress — interim validation numbers, held-out test set untouched.
679k progressions
From-scratch transformer
Beats classical baselines
Live in-browser demo
In progress
L-05
agentic ops · aiops · closed-loop learning
Self-Healing ITOM
An architecture for IT operations that heal themselves. When infrastructure breaks, an agent team detects, diagnoses, remediates from a fixed runbook registry, and independently verifies recovery. A quarantined-reader / privileged-actor split shuts off prompt injection; a policy engine with a four-rung autonomy ladder promotes runbooks on statistical evidence (Wilson lower-bound gating) and demotes them instantly on a failed verify. The core idea: recovery verification doubles as a label source, every verified heal becoming training data for the ML triage prior, a closed learning loop I could not find shipping in any AIOps product. Adversarially audited across eight research lanes; evals centered on false-heal rate.
Diagnose · Remediate · Verify
Recovery-verified labels
Autonomy ladder
False-heal rate
L-06
semantic search · retrieval · 23M papers
Research Field Guide
An independent field guide to ~23 million papers — all of arXiv plus OpenAlex — every abstract embedded into a single 23M-vector HNSW graph and served with hybrid dense-plus-lexical retrieval and cross-encoder reranking, with a no-auth MCP endpoint so agents and LLMs can use it as a tool. Around the search sit ways to actually browse the literature: trending, new and most-cited lists; a zoomable map of research you can play through the years; and a language layer that dates when a term was first attested and who coined it, keeps a dated registry of scientific acronyms, and sorts papers by plainest, most jargon-dense, strangest for their field, and most underrated.
23M papers
No-auth MCP
Hybrid + rerank
Map of research
Concept etymology
L-07
computer vision · embeddings · production
Photo-Selection ML
A production photo system for a matchmaking platform. For each member it scores photo quality and dating-fit, picks the single strongest lead photo, and flags weak sets (near-identical duplicate shots, no full-body, sunglasses, low quality) so members put their best foot forward and the platform keeps its bar high. One CLIP embedding per photo drives the quality models, face-landmark expression and eye-contact features, near-duplicate detection, outfit-variety clustering, and natural-language photo search across ~4,000 photos, served privately and securely for cents a month.
CLIP · MediaPipe
~4,000 photos
Near-dup detection
Semantic photo search
L-08
causal inference · survival · genomics
Dog Longevity ML
End-to-end analysis of the Dog Aging Project, 50,188 dogs, 15,056 deaths, 2,993 variables, genetics for 7,688 dogs. LD-clumped polygenic risk scores, causal forests, TMLE, cognitive-trajectory modeling, kinship GWAS where only IGF1 survives correction, and a novel seasonal-mortality pattern no one had published. All prior DAP work is classical stats only; this is the first modern-ML treatment.
Causal Forests
Survival · RSF
LD-clumped PRS
+760 d Q5−Q1
L-09
genomics · plant pathology · gwas
Net Blotch Genomics
Built with my dad, a plant pathologist: 230 isolates of the barley net-blotch fungus scored against 22 barley lines. A bilinear low-rank compatibility model predicts held-out isolate × line severity at r = 0.90 and estimates ~4 independent effector–receptor systems. 164 of the same isolates matched to public whole-genome sequencing (287K SNPs) for kinship-corrected GWAS and effector presence/absence scans, every claim adversarially audited, negatives kept. In-silico gene pyramiding finds a two-gene stack that stops 99.6% of the panel.
Bilinear MF
r = 0.90 held-out
Kinship LMM
164 genomes · 287K SNPs
L-10
embeddings · semantic search · interpretability
Latent Cinema
Transformer embeddings of 2,390 screenplays, distilled into 50 interpretable "Film Genome" axes. Predicts IMDB rating at R² 0.116 and powers a 1,185-film recommender. Grown into a semantic search engine over 3,418 screenplays embedded scene by scene (394K scenes): search movies by concept, vibe or feeling with the embedding model running in your browser, dial in moods learned from the scripts, and find films that look alike by color palette.
Transformers
PCA
Recommender
394K scenes
In-browser search
L-11
generative · music · autoresearch
classicalML
First large-scale score-level piano dataset (26,934 pieces, 27.5M tokens preserving voice and staff structure) and a 59M-param transformer composer trained via Karpathy-style greedy hyperparameter search. Scaled in 2026 to a from-scratch 755M-param Llama pretrained on 8×H100s over ~21B tokens of piano MIDI: promptable on composer and key, originality-audited against the full corpus (zero verbatim copying), playable in the browser.
PyTorch
755M params
8×H100
L-12
generative image · diffusion finetuning
Image Model Finetuning
Trained LoRA fine-tunes and identity-preserving image-generation models for photorealistic human likeness — capturing a specific person faithfully enough that generated portraits hold up next to real photographs. Entry point into ML; every project since has sat at the seam of ML and creative work.
LoRA
Stable Diffusion
L-13
browser-native data · duckdb-wasm
Spotify Lifetime
An in-browser DuckDB-WASM app over 139K of my own plays (2019–2026), 77+ analytical sections, no backend. Built in a single evening session.
DuckDB-WASM
139K plays
L-14
behavioral psychometrics · llm-as-judge · MCP
personalContext
A pipeline for building a behavioral self-model from one's own communication history. Cross-platform ingest (iMessage · Slack · Gmail · Google Chat) is exposed through an MCP server. An LLM-as-judge scores each relationship against a fixed 10-dimension personality framework. Cross-relational triangulation surfaces the patterns that hold across many relationship types, the higher-confidence signal, and separates them from relationship-specific masks. The pipeline closes the loop by deriving a self-report instrument (yes / no items) that should reproduce the same behavioral profile if administered cold. Behavioral psychometrics, not survey psychometrics. Used self-reflexively only.
MCP server
LLM-as-judge
10-dim framework
Cross-relational
Instrument derivation
