Four builds, each with a published result. Click a tab to open the file, click through for the full case study.
~/the-work/
diffprompt.py
brok.py
cacheguard.py
optiquant.py
$ pip install diffprompt
LIVE ON PYPI
diffprompt
"git diff for your prompt's behavior." LLM-as-judge cascades score behavioral divergence between prompt versions; HDBSCAN clustering surfaces semantic drift so a regression shows up as a cluster shift, not a cosine number nobody trusts.
48tests
3.10–3.12CI matrix
CLI + libCI-ready
$ brok review system.yaml
PUBLIC · LAUNCH-READY
Brok
MCP that reviews system designs with zero LLM in the reasoning path: a 53-entry cited capacity knowledge base, Little's Law estimation, latency/cost lenses, and a 5-rule anti-pattern linter. Abstains instead of guessing.
100/100/0/100/100golden-set
0LLM calls in engine
$ cacheguard bench qqp
OPEN BENCHMARK
CacheGuard
First independent open benchmark of semantic cache false-hit rates. Ships as a drop-in LiteLLM hook with OpenTelemetry spans, using a free local judge that fails safe.
10.5%wrong @ 0.80 thresh
−61.2%wrong hits, after fix
86.9%legit hits kept
$ python backtest.py --live
DEPLOYED · STREAMLIT CLOUD
OptiQuant
LightGBM + CatBoost ensemble with SHAP explainability. Backtested, Dockerised with CI/CD, and actually deployed — not backtested and abandoned.
~/side-quests/
mimic-eval.py
nojargon.ts
ragbait.py
iceman.py
ctxbridge.py
network.ipynb
$ mimic eval --against gpt-4
mimic-eval
Distills expensive LLM-judge decisions into cheap deterministic code — agrees with the real judge ~90-95% of the time. HaluEval kappa 0.96.
$ nojargon scan --tab active
nojargon
Browser extension: on-page due-diligence lens that flags empty buzzwords and tells you what a company actually does.
$ ragbait watch
ragbait
"You opened Instagram again. Enjoy your McDonald's application." Watches your active window, no mercy.
$ iceman predict
iceman
A model that predicts when Drake drops Icemannnnn. Yes, this is a real repo. Yes, it counts as data science.
$ ctxbridge compress
ctxbridge
Compress any LLM conversation into one portable context prompt, pick up exactly where you left off in a new session.
$ python analyze.py
LinkedIn Network Analysis
Growth velocity, company concentration, structural leverage from exported connection data. The 1.5L impressions were not an accident.
