Selected Work
✣HOJB — Research · Product
Lie to Me
"A behavioral analysis system that reads faces, voices, bodies, and words — and caught Lance Armstrong on the first try."
Trained from scratch on the Real-Life Trial Dataset — 121 labeled courtroom testimony clips. Four extractors run in parallel: DeepFace reads facial emotion, librosa extracts pitch and MFCCs, MediaPipe tracks gaze aversion and postural shifts, and Whisper transcribes speech for linguistic pattern analysis.
A Random Forest fusion model learns which combinations of signals predict deception. AUC 0.94 on held-out test data — beating the published baseline of 0.77–0.82 from the original paper. Deployed as a live web app on HuggingFace Spaces.
Out-of-Domain Validation
Tested on Lance Armstrong's doping denial interview — the model flagged it as deceptive via low first-person pronoun use, consistent with Pennebaker's psychological distancing research. The model had never seen footage outside a courtroom.
AUC-ROC — Real Data
0.94
vs published baseline of 0.77–0.82
Top Signal
emo_surprise
Hardest emotion to consciously fake
Training Data
121
Real-Life Trial Dataset courtroom clips
Modality Breakdown — Lance Armstrong
✣HOJB — Personal Project
Wordle.AI
"Two AIs walk into a Wordle. One studied the answers. One figured it out alone."
Live Model Toggle
Built two Wordle solvers using different learning paradigms — one trained on information-theoretically optimal games, one that learned purely from winning and losing. Deployed both behind a single API with a live toggle so you can watch them think differently in real time.
Benchmarked across 2,315 words. The SL model achieves a perfect win rate at 3.46 avg guesses. The RL model reaches 98.2% — trained without seeing a single optimal solution.
Emergent Result
The RL model independently converged to opening with CRANE — the same first guess as the supervised model — without ever being told it was good. It discovered information-optimal play through trial and error alone.
Win Rate — SL Model
100%
All 2,315 Wordle words solved
Win Rate — RL Model
98.2%
No optimal training data used
Opening Word (both models)
CRANE
Emergent — RL converged here independently
✣HOJB — Research
Swahili Emotion Model
"Fine-tuning AfroLM to understand emotion in a language the internet underserves."
Fine-tuned AfroLM for multi-label emotion classification in Swahili, trained on the BRIGHTER dataset (3,307 examples) augmented with cross-lingual English data for a combined 4,621 training samples across 6 emotion classes.
Part of a broader interest in applying ML to problems that matter outside the English-speaking world. The model identifies anger, disgust, fear, joy, sadness, and surprise from Swahili text.
Cross-lingual Training
Combining Swahili with English data via AfroLM's multilingual backbone improved macro F1 from 0.188 to 0.219 — demonstrating that cross-lingual transfer helps even when target-language data exists.
Training Data
4,621
Swahili + English combined
F1-macro
0.219
vs paper best of 0.385
Base Model
AfroLM
bonadossou/afrolm_active_learning
✣HOJB — Research · Systems
OMEGA
"A Transformer that watches crypto, stocks, and forex every hour — and only speaks when it's confident enough to act."
| Pair | Action | ↑ UP | ↓ DN | Conf | Thr | Kelly | Hi-Conf Acc |
|---|---|---|---|---|---|---|---|
| USD/JPY | 🔴 SELL | 47.4% | 52.6% | 52.6% | 50.0% | 2.6% | 68.0% |
| USD/CHF | 🟢 BUY | 56.1% | 43.9% | 56.1% | 54.0% | 6.1% | 61.4% |
| EUR/USD | — HOLD | 51.2% | 48.8% | 51.2% | 55.0% | — | Benched |
Two independent Transformer models trained per market domain — one for crypto and stocks (CryptoCompare + Polygon.io), one for forex pairs (Twelve Data). Both share the same backbone: 3× TransformerEncoderLayer, pre-norm, 4 heads, 24-bar sequence window, softmax over UP/DOWN probabilities.
The system runs autonomously via GitHub Actions every hour: fetch → indicators → scale → infer → threshold → Kelly-size → Discord post → resolve last hour's outcomes → commit logs. The forex model adds temperature scaling and per-pair confidence thresholds calibrated on a strictly chronological validation set.
Design Decision — Silence is a Signal
OMEGA only acts when max(P(UP), P(DOWN)) ≥ threshold. Everything else is HOLD. Most hours it says nothing — that's intentional. EUR/USD and GBP/USD are benched entirely until session-aware features (London/NY overlap) are proven to add edge. Discipline over activity.
Hi-Conf Accuracy — USD/JPY
68.0%
50 signals on chronological test set
Hi-Conf Accuracy — USD/CHF
61.4%
44 signals · threshold 0.54
Position Sizing
Half-Kelly
Capped at 20% of allocated capital
Inference Cadence
Hourly
GitHub Actions · auto-commits outcome logs
