Professional experience
HEARTLAND CENTER · FORWARD DEPLOYED AI ENGINEER
MAY 2025 - PRESENT
Forward-deployed product work: full-stack apps, RAG tooling, and end-to-end ownership for internal teams.
FULL-STACK TIME-TRACKING APP
Built a full-stack time-tracking application in React, Expo, and Tailwind CSS on a Postgres backend, replacing a manual Excel-based workflow. Cut weekly time-entry effort by ~60% across the team.
RAG CHATBOT WITH MCP INTEGRATIONS
Built a RAG-powered internal chatbot with custom MCP server integrations, letting staff query project documents and trigger live workflows in natural language. Cut information-retrieval time by ~50% across the team.
END-TO-END PRODUCT OWNERSHIP
Owned the product experience end-to-end, making the design and UX calls on edge cases like overlapping entries, timezone normalization, and tracking. Drove a ~40% drop in timesheet errors versus the old spreadsheet process.
SQL REPORTING & ANALYTICS
Wrote clean, maintainable code and complex SQL queries powering reporting dashboards and project-level analytics that previously took hours of manual Excel formula work.
POSHDECOR · SOFTWARE ENGINEERING INTERN (FULLSTACK)
SEPTEMBER 2024 - APRIL 2025
Fullstack apps, AI-driven RFP automation, and high-volume data pipelines.
FULL-STACK APPLICATIONS
Built and deployed full-stack applications using React, Node.js, and Python. Lifted system performance ~30% through clean API design and a more scalable service architecture.
DATA PIPELINES & INFRASTRUCTURE
Designed and optimized SQL-based data workflows integrated with AWS, Docker, and Kubernetes. Handled 1M+ data transactions per day across distributed systems.
RFP AUTOMATION PIPELINE
Built an end-to-end RFP automation pipeline using MongoDB, Python, and the ChatGPT API. Processed 80% more proposals per quarter and cut qualification analysis time roughly in half.
PANDAS / PYSPARK REPORTING
Wrote 8 reusable Pandas/PySpark pipelines processing 28M+ rows of source data. Reduced report runtime by ~70% versus the previous batch jobs.
LEAR INC. · ENGINEERING INTERN
MAY 2024 - AUGUST 2024
Asset tracking, low-power hardware, and hardware/software co-design.
END-TO-END ASSET TRACKING
Architected asset-tracking systems spanning RFID, BLE, and computer vision. Drove multi-layer wireless stack decisions across large-scale production deployments.
LOW-POWER HARDWARE DESIGN
Miniaturized multi-sensor hardware through low-power design and tuning. Hit a 3-year battery life target while integrating seamlessly and safely into production assets.
HARDWARE / SOFTWARE CO-DESIGN
Led hardware-software co-design across cross-functional teams from prototyping through deployment. Owned full feature lifecycle and maintained quality across multi-year rollouts.
beSportly · CO-FOUNDER / CTO
MARCH 2022 - MAY 2024
Two-sided booking marketplace for athletes and venues, built solo on the technical side.
TWO-SIDED MARKETPLACE
Launched a two-sided booking platform connecting 460+ venues with athletes. Facilitated 400+ live bookings between players and venue operators.
CROSS-PLATFORM MOBILE APP
Built a single Flutter codebase shipping to both Android and iOS. Collaborated with design to translate Figma mocks into a production app on tight founder timelines.
FIREBASE BACKEND
Implemented the backend on Firebase as a NoSQL server. Tuned data access patterns for quick, low-latency reads/writes against booking and venue collections.
Projects
RETRIEVAL-BRIDGE · PRODUCTION-STYLE RAG RETRIEVAL LAYER
2025
Hybrid vector + BM25 retrieval layer for PromptQL, built on turbopuffer, LanceDB, and pgvector and wired into Hasura DDN.
HYBRID RETRIEVAL ENGINE
Built a hybrid (vector + BM25) retrieval engine for PromptQL, exposed as a search_documents command through a Hasura DDN Python lambda connector. Backed by turbopuffer, LanceDB, and pgvector across the retrieval path.
RANK FUSION & DDN RELATIONSHIPS
Fused semantic and exact-keyword results with reciprocal-rank fusion, then joined hits to Postgres account facts through declarative DDN relationships for grounded, account-aware answers.
OMNO · TIME TRACKING WEB APP (IN PROGRESS)
2025 - PRESENT
Full-stack time tracking app I'm currently building, with a focus on AI-assisted development workflows.
AI-ASSISTED DEVELOPMENT
Frontend prototyping with Vercel v0, then full implementation and iteration through Claude Code. Treating the build as a working sandbox for modern AI-assisted dev practices end-to-end.
FULL-STACK ENGINEERING
Owning the whole stack: frontend, API, persistence, deployment, and iteration. Writing, testing, and debugging in tight loops while defining requirements from scratch and translating them into shipping features.
WHISPERAI SUBTITLE GEN · LOCAL GPU-ACCELERATED TRANSCRIPTION
2024
OpenAI Whisper running locally on my homelab GPU with a Python pipeline for end-to-end subtitle generation.
LOCAL CUDA INFERENCE
Deployed OpenAI Whisper on a homelab server with NVIDIA GPU acceleration via CUDA. Built an end-to-end Python pipeline that yields 95%+ transcription accuracy across multilingual media files.
BATCH DECODING OPTIMIZATION
Tuned inference and batch decoding on CUDA. Cut transcription latency by ~60% relative to CPU-only execution, making large media libraries practical to process locally.
GEAR MESH CV · YOLO DEFECT DETECTION ON GEAR PATTERNS
2024
Computer vision model trained to find mesh-pattern defects on real gear imagery, with strong real-world performance.
YOLO TRAINING & METRICS
Trained a YOLO-based CV model on 7,500 labeled images and hit 88% mAP@0.5. Drove an 80% increase in defect detection rate and reduced false negatives by 21% versus the prior baseline.
JBOD FIRMWARE · LOW-LEVEL C STORAGE CONTROLLER
2024
Firmware for a JBOD storage system written in C: caching, seek optimization, networked control, multithreaded request handling.
STORAGE I/O FIRMWARE
Wrote low-level firmware in C for a JBOD storage system. Implemented disk caching, seek optimization, sequential read/write paths, and garbage cleanup routines, lifting I/O throughput ~3x.
NETWORKED CONTROL PLANE
Built a network access layer for remote JBOD management over TCP. Command parsing, error handling, and concurrent request support via multithreading on the controller side.
HOMELAB · PROXMOX + SELF-HOSTED SERVICES
ONGOING
Personal Proxmox cluster running self-hosted services. Daily-driver sysadmin and Linux experience.
PROXMOX INFRASTRUCTURE
Build and maintain a Proxmox VE server provisioning VMs and LXC containers. Networking, bridges, firewall rules, storage pools, scheduled backups with retention, and capacity planning across Debian/Ubuntu guests.
AUTOMATED IMAGE BACKUP
Linux-based backup system for personal photo/media archives designed against the 3-2-1 principle. Cron + shell scripting for scheduled jobs, integrity verification, and recovery testing.
JELLYFIN MEDIA SERVER
Self-hosted Jellyfin instance reachable remotely. Reverse proxy with SSL/TLS, systemd service management, user access controls, and performance tuning for playback.
NLP · FINE-TUNED FLAN-T5 SUMMARIZATION (CMPSC 497)
2024
Fine-tuned Flan-T5-base on SciTLDR to generate concise approach summaries from scientific papers (AIC text).
DATASET & TOKENIZATION
Loaded the AIC config of SciTLDR (~5.4K papers), filtered prompt/target by length, and split into train/validation/test. Tokenized with the Flan-T5 tokenizer, capping prompts at 128 and targets at 512.
TRAINING & EVAL
Trained Flan-T5-base via HuggingFace Trainer (batch=4, lr=5e-5, 3 epochs, fp16) on a single GPU. Final ROUGE-1 ≈ 0.184; the model exhibited input-copying behavior under the chosen hyperparameters, motivating planned follow-ups around prompt engineering and label smoothing.
ROBOTICS · RTOS-DRIVEN ROBOT CHOREOGRAPHY
2024
Embedded robotics project running a choreographed dance routine on an RTOS, with synchronized motion control, sensor feedback, and audio timing across concurrent tasks.
RTOS TASK SCHEDULING
Built concurrent firmware on a real-time operating system: separate tasks for motion playback, sensor polling, and audio/timing sync. Used queues, mutexes, and priority scheduling to hold deterministic timing across the full routine without dropping beats under load.
CHOREOGRAPHED DANCE ROUTINE
Authored a timeline of servo and motor commands aligned to a song's tempo. Tuned acceleration, dwell, and transitions so the moves read as a single performance rather than a sequence of disjoint steps.
SENSOR-DRIVEN CONTROL
Closed the loop with onboard sensors for stability and timing correction, keeping the routine recoverable when the chassis drifted, encoders fell behind, or actuators stalled mid-step.