Bhagyesh Vora — Applied AI Builds

AI research pipeline · agentic outreach tooling · live web app

• Two-part system: research pipeline producing 100+ source-verified leads across 5 markets and 3 buyer segments, and an outreach tool that drafts client-specific emails and autonomously selects the most relevant collateral per prospect from a master catalog
• Deployed as a live web tool the sales team uses daily, with lead-status tracking and automated weekly digests to the founder
• Est. 100+ manual research hours displaced at $0 in data-license costs

LangGraph supervisor · PostgreSQL (relational + vector + graph) · MCP tool layer · Langfuse observability

• Architected a 10-agent system (ICP discovery → outreach → qualification → proposal) under a LangGraph supervisor — checkpointed state, per-agent token/tool/wall-clock budgets, and policy gates
• Specified HITL autonomy bands (auto-publish → marketing review → sales review): no agent advances a band without a documented eval
• Designed a shadow-mode promotion pipeline — prompt and model changes scored against a golden set before reaching production

Next.js · TypeScript · PostgreSQL/Prisma · Claude Code · Cloudflare

• End-to-end digitization from WhatsApp inquiry to payment — 15-state order lifecycle, tokenized no-login customer flows, immutable quotation PDFs, audit-logged financials
• AI layer built around operator trust: every AI capability lands in a human-approval state, all financials compute as deterministic integer arithmetic independent of model output, and rule-based fallbacks run the full system without live AI credentials
• 90+ tests; AI photography pipeline for a 138-dish catalog; template engine rendering one menu definition into 7 shareable formats

Cloudflare Workers · D1 · Claude API

• Turns session transcripts into longitudinal clinical intelligence — themes, measurement-based-care trends, goal follow-through, auto next-session briefings — cutting documentation and prep ~90 min/session
• Engineered for clinical defensibility: every claim links to transcript evidence, inference is visibly separated from fact, safety-relevant content is flagged rather than interpreted, and nothing enters the record without therapist approval
• 14-threat security model with prompt-injection defenses, audit logging, retention/deletion semantics, and a documented HIPAA gap analysis; 111 tests; written eval plan with rubrics and regression cadence

Demo login available on request

Claude API (Messages + prompt caching) · Chart.js · IndexedDB · Cloudflare Workers

• Six-agent pipeline — statement extractor → merchant intelligence → dashboard narrator ∥ financial analyst → conversational coach + paystub parser — passing typed JSON through shared state
• Accuracy benchmarked against ~2,500 hand-reviewed transactions: ~98% extraction recall, ~90% categorization (85% target), with self-reconciliation against printed statement totals
• Eval harness with a 3-layer discrepancy taxonomy, confidence scores, duplicate signatures, and a human review queue feeding few-shot examples and prompt rules back into the pipeline

cashlens2.pages.dev →

Next.js · TypeScript · Cloudflare Workers/KV · Claude API

• Production web app for a 21-person forecasting league across a 104-event tournament — frontend, edge API, live data pipeline with automatic result resolution, continuous deploy
• LLM insight engine (Claude, tool-use) with anti-hallucination guardrails: the model sees only deterministically computed stats, and every claim is verified against ground truth plus human review before publish