▸Architected end-to-end data pipelines that ingest 1,000+ open-source GitHub repositories and emit SWE-bench-style training instances, automating Docker environment setup, commit-pinned dependency resolution, and isolated test execution for reproducible LLM training and evaluation.
▸Wrote pytest verification harnesses reaching 90–100% line coverage on target issues, producing high-fidelity pass/fail reward signals used to fine-tune agentic coding models at frontier AI labs.
▸Curated and annotated issue, gold-patch, and test training triples across heterogeneous codebases, debugging flaky tests, dependency conflicts, and historical build breakage at scale.
Sep 2022 – Aug 2024
▸Engineered Layer 3 networking stack features in C++ for 5G radio access networks, scaling to support more than 1 million users.
▸Designed system alarms and call performance reports in C that decreased network downtime by 18%, and implemented LTE subscriber redirection that improved handover success rates by 22%.
▸Managed and optimized 15+ developer environments and servers (WRCP, Supermicro), reducing configuration time by 40% and cutting network setup time from 3 days to 4 hours.
▸Created and executed 200+ automated test cases for Qualcomm and Samsung chipset compatibility, increasing test coverage by 35% and identifying 28 critical integration issues before production.
▸Successfully integrated 3 new vendor 5G core systems into existing network architecture, maintaining 99.9% service availability during transition and reducing operational costs by 15%.
Sep 2021 – Jun 2022
▸Designed and programmed C++ firmware for an STM32L072 microcontroller using I2C, UART, GPIO, and ADC interfaces to collect real-time environmental data from seven peripherals.
▸Implemented LoRaWAN communication over the 915 MHz spectrum to transmit sensor data packets from distributed field nodes to a centralized ESP32 gateway across distances of up to 40 meters.
▸Led PCB design across two hardware revisions and successfully deployed the resulting sensor network.
▸Developed machine learning models to analyze sensor-network data for weather-pattern modeling.
Jun 2021 – Sep 2021
▸Built a Django and React B2B web application to provide precise investment analytics to over 10 companies.
▸Implemented Python-based fuzzy search algorithms with over 90% accuracy for financial risk analysis and recommendations.
▸Achieved 99.9% uptime after migrating codebase to an AWS serverless architecture including AWS Lambda, EC2, and S3.
▸Utilized MySQL and AWS Aurora DB for data querying and ran machine learning models with over 90% accuracy to analyze over 100 trends in venture capital revenue and growth.
Sep 2019 – Dec 2019
▸Created Python scripts to transform MySQL data into structured spreadsheet formats for efficient analysis of sports tracking data.
Jun 2018 – Aug 2018
▸Developed a Python web scraping tool that automated data collection for 300+ engineering projects.
Projects
2026
▸Built and trained a 35M-parameter Llama-style autoregressive language model from scratch in PyTorch, implementing multi-head attention, RoPE embeddings, RMSNorm, and SwiGLU activations.
▸Trained a custom 16k-vocabulary BPE tokenizer, completed general pretraining on 8M tokens, and continued domain pretraining on verbalized UFC fight data.
▸The model generalized to held-out prompts and generated syntactically correct, contextually coherent fight summaries.
▸Full-stack UFC pick-em application used by hundreds of users to make predictions, track results, and score points in real time.
▸Built a live scoring engine and idempotent ingestion pipeline that handle fighter replacements, cancellations, weight-class changes, no-contests, and late result corrections without disrupting active picks.
▸Built a model-independent core for prompt-free Mixture-of-Experts routing traces: Switch Transformer hooks, a seven-level compression ladder with bits-per-token accounting, a differentiable routing-camouflage loss, and group-disjoint detector evaluation.
▸Ran a three-seed pilot on google/switch-base-8 on a single RTX 5080: compressed traces reached 0.90 AUROC on the full representation and ~0.74 at ~46 bits per token.
▸Reported a negative result. An untouched model and generic fine-tuning both scored 1.00 AUROC on the controls, so the signal is not backdoor-specific, and named the next two tests needed.
▸Tested whether a model shown its own grader (reward = output % 2) starts optimizing that reward instead of the user's request for an even number, using a three-layer design: a 58-call broad screen, controlled Qwen3-4B interventions, and a 112-row direct-probability probe.
▸Under capped reasoning, Qwen3-4B followed the user in 0/12 mirrored conflicts but 6/6 when explicitly told to maximize reward; the preregistered controllability-by-conflict probability interaction was -0.935 log odds, opposite the reward-optimization prediction.
▸Concluded that for this setup the grader acts as low-authority metadata, not an adopted objective, and argued that diagnosing reward hacking needs causal tracking of the reward, not a single reward-correlated answer.
▸Showed that low address-delta reconstruction loss can still produce failure-scale cache hit-rate errors because cumulative integration breaks cache-block reuse identity.
▸Redesigned the proxy generator around logarithmically binned reuse distance, stateful workload conditioning, and LRU-stack replay.
▸Produced a 565,681-byte deployment package whose largest observed in-sample hit-rate error was 0.0037807 across five evaluated workloads.
▸Developed and programmed internal and external temperature and humidity sensors in C++ to monitor environmental conditions for water generation.
▸Designed and implemented a real-time data logging system using ISRs and RTOS to capture and analyze day–night cycle patterns affecting water yield.
▸Engineered circuit designs to power and control key system components, including cooling fans, power regulation modules, and a water quality sensor for cleanliness monitoring.