I turn models into
business outcomes.

Problem

Travel AI tools hallucinate and can't handle multi-hop constraints like "pet-friendly hotel near a vegan restaurant in a walkable neighborhood." No memory across sessions.

Approach

Adaptive RAG with fine-tuned Mistral 7B query router (94% accuracy, 60ms) selecting between dense, sparse, hybrid, and graph retrieval. BERT personality classifier (7 dimensions, 91% accuracy). Persistent memory for cross-session learning.

Result

Faithfulness: 0.76 → 0.92 on RAGAS. Repeat queries reduced 78%. Constitutional validation catches hallucinations before users see them.

Problem

Paper-based heavy equipment inspections are slow, inconsistent, and can't reference service manuals in the field. Missing defects creates safety liability.

Approach

Full-stack multimodal assistant: photo + voice input, RAG-grounded reasoning with Actian VectorAI retrieving service-manual excerpts, triple-layer hallucination control, schema-validated JSON (Zod) for deterministic UI.

Result

Inspections completable in under 10 minutes. Reliable severity classification with evidence citations. Edge deployment via Cloudflare Workers for fleet-level analysis.

Problem

Product teams spend ~2 hours per experiment analyzing results manually, miss pathologies (SRM, novelty effects, Simpson's paradox), and can't run causal inference without randomization.

Approach

Experimentation platform with Bayesian + frequentist testing, CUPED variance reduction (38% SE decrease), sequential testing, auto-diagnostics for 8 pathologies. Causal toolkit: DiD, Synthetic Controls, RDD, IV.

Result

Analysis time: ~2 hours → under 10 minutes. Validated via 1,000-run A/A simulations (Type I error: 4.8%). Found hidden segment-level effects enabling targeted rollout.

Problem

In federated learning, slower edge clients send stale gradients that poison the global model. FedAsync and FedBuff collapse under high-straggler conditions.

Approach

Momentum-based gradient projection decomposing updates into aligned vs orthogonal components, selectively filtering harmful stale gradients. Quality-aware aggregation with freshness functions and delta-loss scoring.

Result

52–53% accuracy on CIFAR-10 (ResNet-18) with 20–50% delayed clients under non-IID settings, vs 11–28% for baselines. Stable across 1,000+ rounds.

Problem

Crypto analysis requires real-time data from multiple APIs. Manual monitoring is slow and fragmented across tools.

Approach

Autonomous agent with graph-based workflow, GPT-4 reasoning, 3 custom tools for live data. Thread-based memory. Production safeguards: API timeouts, 12K context-window controls, structured logging.

Result

Automated test suite validating tool selection accuracy and multi-tool coordination. Graceful error recovery across edge cases.

Problem

UK property valuation lacks scalable analytics across 25M+ transactions spanning temporal, geographic, and property dimensions.

Approach

End-to-end ML pipeline: ensemble models (RF, XGBoost), SHAP feature analysis, K-means clustering (5 buyer segments, silhouette 0.68), interactive Streamlit dashboards.

Result

RMSE £42K, R² 0.87. Buyer segmentation and seasonal trends enabling strategic investment decisions. Spark-compatible for future scale.

Problem

Job application prep is manually intensive — 30+ minutes per application across JD parsing, resume tailoring, and cover letter writing.

Approach

Two-part system: automated job discovery via n8n workflows, plus GPT-4 resume tailoring and cover letter generation with Pydantic structured output validation.

Result

Reduced application prep time by 70%. Structured outputs ensure consistent quality across hundreds of applications.