AI systems that show their evidence.

Reads owner-approved WhatsApp groups, finds the business events inside them, and exports the company's existing Excel formats.

• The AI never approves business truth. It proposes, staff confirm or correct, and exceptions stay explicit approvals.
• Domain rules are written down. Deal direction, invoice and LPO numbers, AED and USD, tonnes and gallons are always kept with their units.
• Built to run cheaply. Local models were too heavy for my development PC, so it runs on a hosted open-weight model (gpt-oss-120b on Groq).
• Blind testing discipline. A case that exposes a defect leaves the blind set for good, and candidates are frozen before a blind test.

Acceptance target: at least 98% critical-field accuracy on 100 to 200 unseen historical messages, with zero critical hallucinations. This is the target, not a published score.

Watches recorded teaching sessions and scores them on a 14-criteria rubric, with a timestamp behind every piece of feedback.

• Evidence over opinion. Each improvement carries Do, Why, How and the moment in the video it came from.
• Hard signals feed the grader. Talk ratio, interaction, screen share, OCR hits and audio quality are measured first. Punctuality comes from the Excel schedule.
• Weighted rubric. Concept clarity counts most (15%). Balanced activity, differentiation and parent interaction count 10% each.
• Pipeline. FFmpeg, Whisper, camera and screen analysis, OCR, rolling session memory, Qwen2.5 7B grading, HTML, JSON and CSV reports.

Output from the audit system on a test session

Not yet shown: scores come from test sessions. I have not measured agreement with human auditors and would run that comparison before trusting scores at scale.

Photograph an invoice and the data lands in the right account, in the accountant's own Excel format.

• Columns: Serial No, Date, Party, TRN, Description, Amount, Debit to, Credit to. Multi-item invoices become one row per line plus a total row.
• Handwritten routing notes, such as an arrow marked with a company name, decide which account the entry goes to.

Planned flow with sample values. Not live output.

• No third-party AI APIs by design. The logic is built from Python libraries and explicit rules, so it can be owned, self-hosted and audited.
• Missing or blurry fields are flagged for manual entry, not guessed.
• Direction: from data entry toward audit and recovery, with UAE e-invoicing support.

Status: prototype in active development.

Pick an area, developer and project and get a forecast price per sq ft, a conviction score against a 1.0 baseline, and a written verdict.

• Combines Dubai Land Department records, interest-rate curves and news sentiment from transformer models.
• Features include lagged monthly price trends, distance to Metro Blue Line stations and sentiment of tax and visa news.

Output from the Colab dashboard

• Delivered as a Colab notebook and a web dashboard with a competitor scatter plot.

A trading research stack with signal logic kept separate from execution, so each part can be tested alone.

• Mean reversion, moving-average trend following and statistical arbitrage, plus scikit-learn and XGBoost models.
• Walk-forward validation to catch curve-fitting.

Output from a sample run. Not a performance claim.

• ATR-based stops, equity-based position sizing, Sharpe, Sortino and drawdown tracking.

No live-trading performance is claimed.

CFD is too slow to score thousands of candidate wing shapes. This network predicts lift and drag for NACA four-digit airfoils from camber, camber position, thickness, angle of attack, Reynolds number and Mach number.

• 7,205 training points from RANS (SST k–ω), XFOIL and wind-tunnel data.
• Data consistency mattered. Rotating the freestream delayed stall and over-predicted post-stall lift by about 15%, so training data uses geometric twist.
• Activation mattered. ReLU gave non-smooth drag polars. A 6–64–32–16–8–2 tanh network with 3,210 parameters fixed that.
• Tested on unseen shapes. On four withheld airfoils, R² was 0.940 to 0.994 for lift and 0.906 to 0.929 for drag.