Shubhankit Singh

Building an end-to-end research suite — a unified control plane (Tensile) for distributed training, fine-tuning (SFT, DPO, ORPO, GRPO++), inference, and agentic workflows. The core framework powers both internal research and external client engagements.

• Bootstrapped with $200K of personal capital; grew to six-figure revenue through client partnerships.
• Client partnerships: Finance firms, major neo-clouds, and datacenters — helping them train and fine-tune agents with custom SFT pipelines, reward modelling, evaluation harnesses, and inference optimization on the Tensile platform (e.g. Aion, Lanturn, Pavo AI).
• Research: Extremely horizontal — broke down and reverse-engineered all SOTA AI techniques from scratch, spanning distributed systems, post-training, inference, and agents.
• Shipped speculative decoding models (EAGLE3, Arctic speculators for Llama 3.2) and 7 open datasets (44K+ rows, 40+ likes) on HuggingFace.
• Built a SOTA programmatic PDF parser benchmarked against AI and commercial pipelines — 10–20× faster at zero API cost on born-digital formats.
• Public SDKs (10K+ downloads in 3 months): llm-rotate (multi-provider LLM routing), docparser (high-performance document parsing), and gpu-train (remote GPU execution SDK).
• Scaled Research Commons + MathCommons pages to a combined 20K followers.

Building a Behavioural Data Platform with a granular capture layer baked into native desktop applications, automated RL environment generation, synthetic data generation, and end-to-end post-training signal pipelines focused on both the observation and action layer.

• Capture modalities span browser, desktop, CAD / 3-D, and domain-specific application data.
• Generating complete RLHF and long-horizon RL environments with verifiers from expert demonstrations — targeting labs and enterprises.
• Received acqui-hire and licensing offers from 2 of the biggest data vendors in the space.
• As one of Research Commons’ first clients, Lanturn now runs on RC as its infrastructure & research partner — Research Commons owns and manages all of Lanturn’s technology and research, from the capture stack and training infrastructure to post-training and RL research.

• Led Bluetooth HAL work under Silicon Graphics & Media; founding engineer of the Sigma Bluetooth India team.
• Owned engineering for Office Android, Office iOS, and Universal Print integration into M365 mobile.
• Built Networking RAGs to enhance developer productivity; instructor at Microsoft AI School, teaching LLM theory and tooling.
• Authored a proof-of-concept for billions of users running Bluetooth Android applications over Windows via a Linux kernel — covering gaming, hearing aids, and adjacent verticals.

• Trained diffusion models and built social-media AI agents.
• Re-architected the platform: more robust, more secure, AI-ready.
• Drove the B2C → B2B pivot; aligned the product to a steady five-figure MRR.

Built an end-to-end retail trading platform — high-throughput Strategy Engine, Order Management System, data infra, and frontend for a quant-sciences firm focused on LFT algorithms.

• Performance: Numba / JIT compilation and Cython-based optimizations on the critical path; multi-threaded execution for concurrent strategy evaluation and order routing.
• Integrations: Worked with dYdX, Polygon, and Rootstock for exchange connectivity, market-data feeds, and on-chain settlement.
• Shipped Release 1.0; continues to serve on the board.

• Multiple microservices on the seller side — state machines, Kubernetes, ElasticSearch. SCRUM master for the team.
• Expiry Workflow: Activiti (BPMN) workflow making non-expirable invoices expirable — fakes ↓ 10%, seller NPS ↑ 20%.
• Actioning Service: auto-delist and relist on brand or vertical regulation changes — 40 hrs/week of ops bandwidth saved.
• Kubernetes Stateful Onboarding: proof-of-concept for the ElasticSearch Operator on Kubernetes.

• Banking Chatbot: deployed IBM Watson on a Node.js server for NLP and configurable dialogue flow — 24×7 customer service.
• Aadhar Masking (OCR): first-eight-digit masking per new government norms using Tesseract OCR and OpenCV.

• Desktop Capture Agent (Python): Local HTTP server (aiohttp), OS-level input & screenshot capture (pynput, mss), network monitor, SQLite persistence, pywebview dashboard, pause/resume, batch REST import to cloud. Packaged via PyInstaller (DMG / EXE).
• Plugin Monorepo (v2.7): Unified event schemas; plugins for Chrome/Edge/Firefox (MV3 extension, React 19), VS Code (lanturn-code-capture), Excel (xlwings + Office.js), CAD (AutoCAD/PyRx, Siemens NX, SolidWorks C# add-in).
• Cloud Backend: FastAPI + SQLModel + PostgreSQL microservices — API gateway, browser & desktop ingestion, event & screenshot writers, workflow workers (LLM-powered). Redis, GCP Pub/Sub, GCS, Qdrant; Alembic migrations, OpenTelemetry, Sentry.
• Dashboard (Next.js 15 / React 19): Operator UI for sessions, mined workflows, training datasets, eval rubrics, data evaluations, campaigns, org management. TanStack Query, Tailwind 4, React Aria, Recharts.
• Data & Eval: Short-horizon and long-horizon training dataset pipelines; golden-reference, pairwise, and score-based eval rubrics; automated RL environment generation from expert demonstrations.
• Stack — Python, TypeScript, Go, C#, React 19, Next.js 15, FastAPI, PostgreSQL, Redis, GCP Pub/Sub, GCS, Qdrant, PyInstaller, esbuild.

• Control Plane: FastAPI + Temporal + PostgreSQL service managing multi-cluster job orchestration, real-time WebSocket streaming (LISTEN/NOTIFY), and multi-cloud provisioning (GKE, Nebius).
• Cluster Agents (Go): Custom Kubernetes operator (TrainingJob CRD → Volcano gang scheduling), cluster-watcher (informer → Postgres NOTIFY), RBAC controller (team-scoped isolation).
• Tensile-Train: SFT / DPO / ORPO / RL (PPO, GRPO++) pipelines with async checkpointing, Nydus image acceleration, preemption recovery, and heartbeat liveness.
• Tensile-Infer: Multi-engine serving (vLLM, SGLang, TensorRT) with KEDA autoscaling, speculative decoding, disaggregated prefill/decode.
• Tensile-Agents: Financial, browser/computer-use, and API-MCP agents; modular enterprise RAG with RAGBench evaluation.
• Stack — Python, Go, TypeScript, React 19, PyTorch, DeepSpeed, FSDP, Volcano, Helm, PostgreSQL, Temporal.

• Mathematics education platform under Research Commons with full product, GTM, design, and community strategy.
• Part of the combined 20K followers community footprint.

• End-to-end pipeline: Azkaban → Python → FastAPI → GCS → BigQuery external tables → Grafana.
• Cost-aware migration from Postgres; partitioned JSON contract with on-call playbook.

• cpptensor: C++ tensor operations library.
• cppgrad: C++ autograd engine with reverse-mode AD.
• cppnet: C++ neural network library built on top.

• PDE solver for exact European put / call pricing under Black-Scholes; perpetual American options; exact solutions for the Greeks.
• Monte-Carlo simulations approximating put / call prices and their reactions to expiry and simulation count.
• Compared exact solutions against advanced finite-difference and finite-element methods.
• Stack — C++, Boost, Black-Scholes, Monte Carlo, FDM, FEM, stochastic calculus.

• MATLAB compiler in C++ with a static code analyzer built on unordered maps, equivalence relations, and various hashing techniques.
• Stack — C++, Boost, dynamic and structural equivalence, compiler construction.

• Merit-based admission offered by Worldquant, a quantitative hedge fund. 100% tuition-free.
• Coursework: Financial Markets, Econometrics, Derivative Pricing, Stochastic Modelling, ML in Finance, Portfolio & Wealth Management.

• Graduated with First Division.