I build LLM applications. At DeepLearning.AI I am the founding and sole engineer on three AI recruiting products: an AItechnicalinterviewerusedinlivehiring, a recruitersourcingandreviewtool, and a candidate-facingcareeragent. Before that I built data and AI tools at Bexorg and Tassel, and flight-path analysis at NASA Ames.
Building now
Career agentBuildinghelpscandidatesfindtheirnextroleandreturnsagrowthreport
Built the AI technical interviewer used in live hiring. It scores candidates on what they say, and each score shows the quote it came from, so a recruiter can check it.
Built and maintain the recruiter sourcing and review tool: migrated it from Railway to AWS with zero-downtime deploys, added bulk actions for 1,000 candidates, and cut LLM scoring costs.
Designed and built the Valence Career Agent, which helps candidates find their next role: it learns their goal, background and job constraints, runs a technical assessment, and returns a growth report with a positioning statement, answer coaching and a three-part growth plan.
Built a job-data pipeline tracking careers pages at 1,400+ companies, with 99,000+ jobs collected and 78,000+ live, to power role recommendations.
Exposed the sourcing tool as a scoped API for internal tools, and published dlai-mcp-server on npm for searching DeepLearning.AI courses from Claude Code, Codex and Cursor.
Projects 4AI Technical Interviewer · Recruiter Sourcing and Review Tool · Valence Career Agent · dlai-mcp-server
Projects 4
AI Technical Interviewer
Structured technical interviews, graded from the transcript.
Recruiter Sourcing and Review Tool
Candidate sourcing, scoring and review for the recruiting team.
Moves up to 1,000 candidates through a hiring stage at once with bulk actions.
Serves other internal tools through a scoped API.
Valence Career Agent
Shows job seekers where they stand and what to work on next.
Starts from the candidate's goal (finding their next job, navigating their career, improving their AI skills, or preparing for interviews) and adapts the conversation to it.
Imports their résumé, LinkedIn and GitHub for them to review, then asks only what it does not know: career history, projects, pay, timing, location, visa constraints and non-negotiables.
Returns a growth report that coaches two or three answers (what worked, what to add) and sets three growth areas, each with an action, a deliverable and a success signal, all cited from the transcript.
Supports real-time voice and dictation, and lets candidates own their career snapshot and conversation record, with export and deletion.
Turns a tech blog post into a podcast: content extraction, script generation and text-to-speech.
Crawls the engineering blogs of companies such as Uber, Meta, Netflix and Airbnb on a daily schedule (APScheduler).
Extracts content with several strategies (trafilatura, with a BeautifulSoup fallback) and scores its quality.
Writes a source-grounded summary and a two-speaker script with Ollama Cloud, then voices it with Kokoro on CPU.
First version: GPT-5.2 scripts and OpenAI TTS, with parallel TTS generation in AsyncIO, 3-5x faster for 20-segment podcasts.
Takes user submissions: paste any URL or text and get a podcast in minutes.
Rebuilt the frontend in one day across 28 PRs: auth, player, navigation, library, search and PWA.
Works as a PWA with offline support, lock-screen controls (MediaSession API), iOS safe areas, dvh viewport, WCAG 2.5.8 touch targets, maskable icons and a service worker.
Covers the app with 34 E2E tests across 4 device profiles.
Built a video pipeline that segments graduation ceremony recordings by student and delivers personalized clips to 2M+ students, an $8M+ revenue opportunity.
Built a multi-agent transliteration chain on OpenAI that converts each student's name into its native-language script, so an ElevenLabs native voice pronounces it correctly in ceremony announcements.
Moved heavy Azure pipelines from always-on to on-demand scaling, cutting costs 40% and processing 8x faster.
Projects 2Multilingual Name Pronunciation · Graduation Video Pipeline
Projects 2
Multilingual Name Pronunciation
Getting every graduate's name right in ceremony announcements, in any language.
Closed the gap left by ElevenLabs text-to-speech, which on its own mispronounced many students' names from other languages.
Graduation Video Pipeline
Personalized graduation ceremony clips at scale.
Architected the Azure infrastructure behind it.
StackAzure ·Python ·OpenAI ·ElevenLabs ·FFmpeg ·Voice AI ·TTS ·Video Processing
May 2024 – May 2025 ·New Haven, CT
Bexorg
May 2024 – May 2025
Full Stack AI Engineer · New Haven, CT
Applied NLP extraction and computer vision for medical imaging, with ChromaDB and the Gemini API.
Built a natural-language analytics tool that turns plain-English questions into Plotly and D3 dashboards, cutting analysis time 75%; a key sales differentiator that helped convert pilots to contracts.
Built a real-time data visualization platform on Neo4j and BigQuery with role-based access for 40+ users, cutting retrieval time 45%.
Built an ML quality-control system for time-series data with real-time anomaly detection, reducing data issues 70%.
Resolved critical pipeline bottlenecks, improving throughput 10x and enabling new enterprise customers.
Projects 4GenAI Query Canvas · Interactive Data Viewer · Time Series Analysis & QC · Team Collaboration Calendar
Projects 4
GenAI Query Canvas
Natural language to visualization, with GenAI.
Built automatic chart selection and customization in Plotly and D3.
Added feedback-based learning that improved query accuracy 40% over time.
Interactive Data Viewer
Real-time data visualization on Plotly, Flask, Neo4j Aura and BigQuery.
Built graph visualization in vis.js for exploring complex relationships in real time.
Built a graph-based unified data access layer with GNNs and optimized graph algorithms.
Designed interactive dashboards with real-time updates, improving data analysis efficiency 60%.
Added audit logging to the role-based access, for users across departments.
Time Series Analysis & QC
Time-series analysis and quality control.
Built a visualization engine for time-series data that sped up pattern identification 55%.
Routed automated alerts by severity, cutting response time 80%.
Created quantitative QC scores for quality assessment and trend analysis.
Team Collaboration Calendar
Async communication and experiment scheduling.
Built an async-first calendar that cut experiment deployment time from 7 days to 3-4 hours.
Built an automated scheduling engine that cut 1:1 meetings 60%.
Added real-time experiment tracking with automated status updates and notifications.
Integrated Slack and Jira for cross-functional teams.
Analyzed flight-path adherence to air traffic control procedures with GeoPandas and Pyproj, identifying navigation patterns that informed aviation fuel-cost optimization.
Automated data validation across 1,000+ flight paths, reducing manual verification from 4 hours to 15 minutes, and containerized the legacy analysis system.
Projects 1Flight Path Optimization System
Projects 1
Flight Path Optimization System
Flight-path analysis and adherence checks with Python, Pyproj, Xmitree and Folium.
Converted flight paths from lat/long to Cartesian coordinates with Pyproj, for web maps in Folium.
Built interactive visualization dashboards for airspace data in Folium, Leaflet.js and Plotly.
Containerized the legacy system, cutting setup for new researchers from days to minutes.
Built a time-series view of each flight path for real-time adherence analysis.
MCP server making DeepLearning.AI's 121-course catalog searchable from Claude Code, Codex, and Cursor. 3 tools: search_courses, get_course_details, list_topics. Published on npm, hosted on Railway.
Turns engineering blogs from 15 companies into daily AI-narrated podcasts. Ollama Cloud writes a source-grounded summary and a two-speaker script; Kokoro voices it on CPU. Next.js player, FastAPI and SQLite backend.
Multiplayer middleware for AI coding agents: two developers share one Claude Code session through a Telegram bot. Pair and solo modes, GitHub issues to claim, driver handoff with a context summary, auto-commit and a pull request at the end.