Selected work
Client delivery and AI implementation
The problem, what I built, and the current state of each project. Demos and public code are linked where available.
Live demo
Preparing for commercial launch
The problem: contractor discovery and procurement depend on fragmented records, unverified claims, and bid processes that are difficult to compare. I built a bilingual product for a founder that brings discovery, verified profiles, project posting, comparable bids, awards, milestones, purchase orders, invoices, and audit records into one workflow.
My work: translated procurement requirements into workflows, permissions, application logic and English and Arabic browser cases. Built the demo and prepared the product for client review.
Current boundary: the product and bilingual state coverage are working. External verification, messaging, and payment services remain behind replaceable adapters for production integration.
Partner backed product
LaserEyedBen
Production system
Paused after operation
A crypto intelligence agent that combined market, blockchain, developer, news and social signals. I implemented LLM query classification and routing between PostgreSQL and Weaviate for transactional records and semantic retrieval, alongside external data integrations and failure handling.
Public evidence: a case study and selected implementation samples. The repository is not the complete production service.
Current boundary: the system ran publicly and was paused when API and infrastructure costs outgrew available funding.
Working prototype
Evaluation system
A product specification review system for missing acceptance criteria, conflicts and unresolved decisions. I built an Anthropic model adapter, Zod schema validation and bounded revision loops. Proposed edits stay separate from the source document, and judgment decisions remain with the reviewer.
Working prototype with automated application tests. The visible demo uses a deterministic stand-in when no model key is present; passing those tests does not measure live model accuracy.
Repository available on request
A Python and FastAPI research application with custom agentic orchestration: research framing, query generation, external search, streamed report synthesis and citations. I implemented query fallbacks and checks for report structure and numeric source grounding, plus document export workflows.
MVP with source-grounded generation and heuristic quality checks. Citations are not a guarantee of accuracy; live model evaluation and quality failure handling need further work.
A decision-quality journal for thesis-driven investors. A thesis cannot save without a falsifier, versions remain immutable, reality checks flag broken assumptions, and a deterministic process score evaluates the quality of the decision rather than the return.
No recommendations, broker connections, or automated trading. The product is deliberately limited to recording, checking, and reviewing a user's own reasoning.
