I build AI systems end to end (fine-tuning, agents, computer vision and deployment) and evaluate each one against a baseline on held-out data.
01 Legal Contract LLM · QLoRA fine-tuning with before/after evaluation
Fine-tuned Qwen2.5-7B with QLoRA (4-bit, 0.5% of weights trained) on CUAD — 510 expert-annotated contracts — for clause classification and grounded clause Q&A. Trained on a free Kaggle T4; served on a 6 GB laptop GPU.
| Held-out set (400 items) | Zero-shot | 3-shot | Fine-tuned |
|---|---|---|---|
| Classification accuracy | 61.5% | 63.5% | 80.0% |
| Missed clauses | 57% | — | 5% |
| Calibration error | 0.28 | — | 0.03 |
| Hallucination rate | 1.0% | — | 3.5% |
Deterministic harness, no LLM judge, hallucinations audited by hand. The hallucination increase is the trade-off, reported rather than hidden.
PyTorch Transformers PEFT TRL FastAPI Streamlit → Repository · Model · Video
02 Adversarial Verification System · multi-agent checking with an evaluation harness
A Generator → Critic → Coordinator pipeline in LangGraph that checks LLM outputs against retrieved evidence (hybrid dense + BM25) before returning them, with calibrated confidence scoring.
| 300 evaluations (50 prompts × 2 batches × 3 runs) | Baseline | Verified |
|---|---|---|
| Hallucination rate | 4.0% | 0.7% |
| Accuracy | 76.7% | 86.7% |
| False-escalation rate | — | 2.7%, every run |
LangGraph OpenAI API ChromaDB FastAPI Docker MLflow → Repository · Video
03 Multimodal Behavioral Intelligence · M.Sc. thesis
Fuses vision, speech and language to analyse behavioural consistency: trust-aware multimodal fusion, temporal modelling and explainable reasoning, deployed end to end on FastAPI.
flowchart LR
V[Video] --> M[MobileNetV2] --> L[BiLSTM]
A[Audio] --> W[Whisper] --> D[DistilBERT]
L --> F[Trust-aware fusion]
D --> F --> X[Explainable reasoning]
70.83% accuracy · 81.34% precision · 70.41% F1
PyTorch MobileNetV2 DistilBERT BiLSTM Whisper OpenCV Docker → Repository · Demo video
04 Sensitive-Area Intrusion Detection · real-time computer vision
A fine-tuned YOLOv8 detector with DeepSORT tracking and a rule engine for zone-intrusion, weapon and dropped-object alerts. Incident-based alerting (one event, one alert) streamed live to a React front end over WebSockets.
87.8% precision · 79.7% recall · 86.0% mAP50
PyTorch YOLOv8 DeepSORT FastAPI React Docker DVC → Repository · Try it · Video
Toolkit
Also: Hugging Face (Transformers, PEFT, TRL) · LangGraph · OpenAI API · ChromaDB · MLflow · DVC · YOLO · Whisper · C / C++
