Recent work, deeper cuts

Problem

Health insurance support teams drown in repetitive queries — agents waste time on FAQs while complex issues get delayed and customers never hear back reliably.

Approach

Built an AI agent (Claude Haiku 4.5) with a single `smart_search` tool that first hits a pgvector knowledge-base (with an LLM judge to verify ≥60% relevance), falls back to Tavily web search scoped to admin-approved domains, and escalates to a ticket flow that sends confirmation + resolution emails via Resend. Resolved tickets feed back into the KB.

Value

Customers get instant, accurate answers for known topics; complex issues are escalated with a full audit trail; every resolved ticket makes the system smarter over time.

Problem

Manual peer review is slow, inconsistent, and doesn't scale — reviewers apply different rubrics, miss key dimensions, and create bottlenecks.

Approach

Built a multi-agent pipeline with two sequential waves: Wave 1 runs Grammar, Novelty (with live Google Search), and Fact-Check agents concurrently; Wave 2 runs Consistency and Authenticity agents in parallel; a final Evaluator agent applies weighted scoring to produce a PASS/FAIL verdict.

Value

Structured, explainable peer-review reports in seconds — traceable per-dimension scores with rationale, not just a verdict.

Problem

Debugging across large codebases is slow — devs context-switch between docs, search, and code without a unified interface that understands the project.

Approach

Built a full MERN stack tool with an ingestion pipeline (POST /ingest) that indexes codebases into a vector store, a search layer (GET /search) with filters by username, project, and producer, and a RAG endpoint (POST /ai) that runs tool-call to vector search and returns an AI-generated answer.

Value

A project-aware debugging assistant that answers questions grounded in the actual codebase — not generic docs.

Problem

GenAI apps break when providers fail, outputs drift, and retries aren't standardized.

Approach

Designed a unified interface with routing, retries, and fallbacks built-in.

Value

Faster experimentation, less integration glue, and more reliable production behavior.

Problem

Handwritten prescriptions are hard to read and error-prone.

Approach

Used a vision-capable model with a web workflow for clean, readable output.

Value

Improves clarity and reduces errors for patients and pharmacies.

Problem

DNS visibility and updates were fragmented across tools.

Approach

Built a single dashboard with MERN and AWS SDK integrations.

Value

Gives teams a clear, centralized view of DNS state.

Problem

Needed a clean retail UI baseline.

Approach

Built a responsive landing page with Tailwind and React.

Value

Strong baseline for retail product presentation.

Problem

Needed model-driven insights from noisy real-world data.

Approach

Built pipelines and models for sentiment and stock prediction.

Value

Demonstrated end-to-end ML modeling and evaluation skills.

Problem

Explored event-driven UI and gameplay mechanics.

Approach

Built a lightweight game with vanilla JS and DOM updates.

Value

Showcases interaction design and performance on the web.

Problem

Language apps rely on tapping buttons and filling blanks — learners never actually speak, so they can't build real conversation skills.

Approach

Built a voice-first agent with Deepgram STT → GPT-4o tool dispatch → Cartesia TTS, running over LiveKit WebRTC. The agent switches between 5 modes (casual chat, structured lessons, quizzes, roleplay conversation, doubt resolution), semantically grades answers via LLM, persists sessions to Neon Postgres, and logs per-turn latency metrics.

Value

Learners practice real spoken conversation, not screen-tapping. Every turn feels like talking to a real tutor — and the system gets smarter about each learner's progress.