Tejas Singh Bhati CS undergraduate, IIIT Sonepat
Open to internships 00:00:00
Pieter Bruegel the Elder, Hunters in the Snow, 1565
Plate 01 / P. Bruegel the Elder / Hunters in the Snow / 1565 / Oil on panel, 117 x 162 cm
Tejas Singh Bhati

Tejas Singh Bhati

Third year CS undergraduate at IIIT Sonepat. I work on infrastructure and scientific computing: node readiness in Kubernetes, the integration layer between the UN's climate and macroeconomic models, and a code editor that runs models locally. I build my own tools too, usually to find out how something works.

Programme
B.Tech Computer Science, Data Science and Analytics
Institute
Indian Institute of Information Technology Sonepat
Year
Third year of four, 2024 to 2028. CGPA 8.77
Coursework
Data Structures, Algorithms, Operating Systems, Database Systems, Computer Networks, Artificial Intelligence, Machine Learning
Interests
Distributed systems, model integration, agentic runtimes
Location
Sonepat, India. Working remote
Contact GitHub LinkedIn YouTube
01 / Experience
2026.08 to present Remote

Cloud Native Computing Foundationkubernetes-sigs/node-readiness-controller

Open source contributor

  • Corrected a directional comparison in the node update predicate. Label state was diffed in one direction only, so removals never enqueued a reconcile and readiness gating went stale.
  • Extended the manager ClusterRole to cover events.k8s.io, which the packaged RBAC omitted, blocking event emission under the shipped Helm chart.
  • Removed an empty nodeSelector from the chart's rendered output, which was propagating an invalid manifest suggestion to downstream operators.
3 mergedGo, controller-runtimeApache-2.0
GoKubernetescontroller-runtimeHelmRBAC
2026.07 to present Remote, Noida

SuperfluenticFounding engineer

AI automation and full stack

  • Build workflow automation on LLM APIs, wiring model inference into business logic and internal tooling rather than shipping standalone assistants.
  • Work the full path from interface to deployment: Next.js and React on the client, Node and Python services behind them.
  • Contribute to early product definition and iteration as a founding engineer, including the public landing surface and its rebuild.
Founding engineerNext.js, Node, PythonLLM APIs
Next.jsReactNode.jsPythonLLM APIsProduct
2026.05 to present Remote, Boston

Mnemosyne SystemsProject orangu

Major contributor

  • Own the workspace subsystem end to end: workspace model and manager, in-place directory switching, live tab activation under Alt bindings, and stream-completion signalling on parked tabs.
  • Implemented local semantic code search over hybrid retrieval, 1,612 lines across 20 files, with the index held outside the workspace tree.
  • Shipped twelve tools including /schedule for cron-style scheduled commands, plus /bisect, /stash, /branch, /restore and /pull_request.
  • Added UUID-backed sessions with branch-aware auto-resume and per-workspace LLM configuration, isolating model and server selection per tab.
  • Built the automated documentation pipeline on GitHub Actions and maintain the project site across nine merged pull requests, including release pages and PAT-authenticated publishing.
35 merged12 tools shipped6,000+ lines added
PythonCLI architectureGit internalsHybrid retrievalGitHub Actions
2026.02 to 2026.04 Remote, New York

United Nations OICTEAPD-DRB, UN DESA

Open source contributor, scientific programming and model integration

  • MUIOGO couples climate, land, energy and water resource models to a dynamic macroeconomic general equilibrium core across ten country configurations. I work on the service and integration layer.
  • Hardened the Flask API: session ownership verification ahead of destructive case operations, structured JSON error responses in place of bare OSError propagation, and file handle leaks in FileClass closed by moving to context managers.
  • Resolved silent configuration corruption in the OSeMOSYS parameter build, where a duplicate RYTM key across DEFAULT_F, UPDATE_F and GEN_F overwrote solver settings, and added null guards through the parameter dictionary construction.
  • Authored OG-CLEWS: Linking of Linear Resource Optimization and Non-Linear General Equilibrium Models, identifying seven structural integration frictions and benchmarking each against the decomposition strategies used in MESSAGE-MACRO and TIMES-MACRO.
  • Built the integration prototype: OG-Core runner, bidirectional ETL across resource and economic schemas, L2-norm convergence loop with structured logging, and a validated renewable transition scenario for Mauritius.
14 merged1st of all GSoC 2026 applicants10 country configs
PythonFlaskOG-CoreOSeMOSYSETLNumerical convergence
02 / Upstream

Pull requests merged into public repositories owned by other people, read directly from the GitHub API with diff statistics. Private and client work is excluded.

03 / Activity

0 contributions in the trailing twelve months.

LessMore
04 / Projects

MARS

Active

Modular agent runtime system

Runtime for long-lived, tool-using language agents under realistic operational constraints. Six independently deployable modules: persistent storage, multi-platform message normalisation, hot-loadable plugin execution, LLM orchestration, dual-layer episodic and semantic memory, and a proactive scheduler.

Node.jsSQLiteChromaDBRAG

nrc-explorer

Active

Headlamp plugin, Kubernetes

Resolves node admission state in the dashboard: readiness conditions, applied taints and reconciliation history for a single node, surfaced together rather than reconstructed from controller logs.

GoKubernetesTypeScript

Drishya

Shipped

Geospatial asset detection, Indian Railways

Detects and classifies urban assets from satellite, aerial and drone imagery through a hybrid pipeline: YOLOv8 for discrete objects, spectral segmentation for vegetation, water and surface classes that carry no reliable edges.

FastAPIYOLOv8ReactDocker

Kaavish

Active

Automated adversarial testing for AI systems

Executes prompt injection, jailbreak, data exfiltration and agent hijacking classes against deployed agents and LLM applications, returning reproducible proof-of-concept transcripts rather than severity scores.

PythonLLM securityRed teaming

pathosim

Shipped

Stochastic epidemic simulator

Compartmental disease dynamics over a D3 world map with configurable basic reproduction number, case fatality rate, incubation period and transmission mode, propagating across countries with per-milestone event generation.

JavaScriptD3.jsStochastic modelling

OEECF

Shipped

Epidemiological and economic coupling framework

Translates epidemiological state into labour supply and productivity shocks, giving health and macroeconomic models a shared unit basis so policy scenarios can be evaluated across both simultaneously.

PythonCoupled modelsMacroeconomics
05 / Stack
06 / Awards