Work

Experience across startups, research labs, and academia.

✦

RL Residency

Prime Intellect · 03/2026–05/2026 · San Francisco, USA

Undertow

  • Built Undertow, a synthetic PTC benchmark with hidden mutable worlds and dependency graphs.
  • Designed hidden_dag and repair_chain task families to test value ledgers, tool sequencing, stale observations, repeated calls, constraint failures, and runtime budgeting.
  • Led PTC vs native tool-use evaluations on Prime Intellect infrastructure, using trace analysis and budget calibration to separate true progress from repeated-call loops, stale-state errors, and timeout behavior.

Sparse Signal Loop

  • Designed and ran Sparse Signal Loop, a controlled 2×2 study of chat vs RLM and dense vs sparse judge feedback across LongBench-Pro and Mini SWE Agent Plus.
  • Extended with notebook/file persistence and skill-file variants to test how feedback sparsity and memory placement changed refinement behavior.
  • Showed that management choices mattered more than heavier harnesses or durable file memory, with judge agreement sometimes diverging from real task success.

Environments & LongCoT Evals

  • Built a broad suite of RL environments — Factory Sokoban, GraphWalks, ritual-machine, long-context-retrieval, LHAW, LOCA-bench, ProgramBench, and rubric-discovery — with paired baseline and PTC/standard variants for controlled ablations.
  • Implemented LongCoT / LongCoT-RLM environments and ran hosted evaluations showing strong RLM gains on long-horizon reasoning, including Qwen 3.5 9B reaching 15.69% on full LongCoT and Qwen 3.5 27B reaching 22.18%.

Founding AI Engineer

QX Labs · 05/2025–10/2025 · London, UK
  • Architected and built a production agentic knowledge-graph for private-equity workflows supporting multi-hop natural-language queries, streaming responses and multi-database storage (graph, vector, NoSQL).
  • Designed an agent orchestration layer for query routing, async execution and optimized database query generation for multi-hop queries.
  • Built scalable ETL and ingestion pipelines with ~3 million company records from NoSQL into Neo4j, for periodic updates to the knowledge graph.
  • Developed graph-based recommendation algorithm (attribute bucketing + weighted seed expansion); reduced recommendation & query latency from 42s to 10–12s.

Founding AI Engineer

Unsiloed AI (YC F25) · 01/2025–07/2025 · San Francisco, CA
  • Built extraction, classifier, splitter APIs, document chunking strategies, all with streaming support for Fortune 500 firms.
  • Fine-tuned vision models (YOLOv11) for document-layout segmentation (10 classes) and optimized inference across OpenVINO / TensorRT / ONNX runtimes for CPU/GPU deployment.
  • Trained and benchmarked field-extraction models on 10k+ financial reports and SEC filings; developed an evaluation suite and achieved top performance vs competitors.
  • Implemented spreadsheet-encoding techniques for robust LLM-ready extraction (merged cells, cap tables) and a human-in-the-loop pipeline to continuously improve labels and model quality.

Tech Lead

Covenants PharmaChem · 10/2024–07/2025
  • Building an AI platform for end-to-end procurement automation.

Research Mentee (independent work)

Independent research · Tsinghua University · 09/2024–01/2025
  • Replicated and extended parts of the Canonical Representation Hypothesis (CRH) work (delayed generalization + representation dynamics).
  • Tsinghua: Informal collaboration on literature reviews in 3D reconstruction and generation.

Research Intern

IIT Bombay · 08/2024–01/2025
  • Spin-Neuro-Quantum-AI Group: Developed and implemented spiking neural networks for German Speech classification. Tech stack: Python, snnTorch, PyTorch, NumPy, Pandas.
  • Real-Time Early Fault Detection: Built a real-time fault detection system predicting faults in electromagnetic machines via digital signal and magnetic flux analysis. Tech stack: Python, PyTorch, Librosa, SciPy, Pandas, DSP.

Teaching Assistant, Advanced Algorithms and Complexity

Bachelor's — Computer Engineering · 09/2024–12/2024
  • Led problem-solving sessions and mentored 16 master's students, enhancing their grasp of advanced algorithms.
  • Designed problem sets based on research papers and lectures to promote critical thinking beyond the curriculum.

Coordinator

IEEE CS Student Chapter · 11/2022–12/2023
  • Coordinated events, meetings, conferences, hackathons, and competitions to ensure smooth operations.

Developer Analyst Intern

Predoole Analytics · 06/2023–08/2023
  • Built an email-generation API for personalized outreach; integrated Apollo.io data-fetch APIs.
  • Tech stack: Flask, LangChain, OpenAI API, Python.
✦

“Game Is Game.”