Suraj Ranganath

selected work

• BalkanID Copilot production AI agent for identity security over enterprise identity graphs

  Helped build an agentic identity-security product used by enterprise customers. The useful part was not the chat box; it was the graph, permissions, evals, monitoring, and action boundaries around it.

  • Text-to-Cypher over a multi-tenant Identity Intelligence Graph.
  • MCP-style action flows, connector workflows, safe rollout loops.
  • 6+ enterprise customers, $200K+ product revenue, 4 patent filings.

  agentsknowledge graphsIAMMCP
  docspatent
• Show, Don’t TELL explainable AI-generated text detection with evidence spans

  The detector does not just produce a score. It marks the spans it is reacting to, which makes the interface easier to inspect and argue with.

  • Newest paper currently indexed for my scholarly profile.
  • Built around the same detector robustness thread as StealthRL.
  • Supported by merged work in the public ACMCMC/rl-detector repo.

  interpretabilitydetectionevidence spans
  paperdemogithub
• StealthRL fairness-aware RL red-team loop for AI-text detectors

  GRPO + LoRA paraphrase training against a detector ensemble, with rewards for evasion, semantic preservation, fluency, and ESL fairness pressure.

  • Uses Qwen3-4B-Instruct adapters and detector-transfer evaluation.
  • Accepted for poster presentation at Amazon Trusted AI Symposium 2026.
  • Public repo currently leads my GitHub by stars.

  AI safetyGRPOLoRAfairness
  paperdemogithub
• EEGPrep Python port of EEGLAB, the popular toolbox widely used by EEG researchers

  Current open-source work at the Swartz Center on a Python port of EEGLAB, the popular toolbox widely used by EEG researchers. Contributions span GUI parity, rejection/component workflows, plotting, DIPFIT, STUDY metadata, clustering, statistics, packaging, and docs.

  • 30 merged PRs in sccn/eegprep as of May 30, 2026.
  • Additional SCCN PRs in eeglab and eeglab_conversion.

  open sourceEEGPythonresearch tooling
  githubprs
• NAMO and NAMO-D in Marin implemented two new optimizers for foundation models in Marin, mentored by Al Merose

  Implemented NAMO and NAMO-D optimizer support for foundation-model training in Marin/Levanter with mentorship from Al Merose: optimizer logic, config classes, registry wiring, exports, tests, and clamping validation.

  open sourceoptimizersJAXMarin
  github
• KV Cache Quantization for Self-Forcing Autoregressive Video Generation 33-method study for self-forcing autoregressive video generation

  A systems-heavy study of how far KV-cache compression can go before self-forcing autoregressive video generation starts drifting or losing useful quality.

  • Tracks VRAM, runtime, compression, VBench, SSIM, LPIPS, PSNR, and drift.
  • Finds practical INT4 soft-prune regions around 5.4× compression.

  world modelsvideo diffusionquantizationsystems
  papergithub
• PIPE-RDF LLM-assisted pipeline for enterprise RDF benchmark generation

  Public KGQA datasets rarely match private enterprise schemas. PIPE-RDF builds schema-specific NL-SPARQL benchmarks with validation and repair.

  • Reverse queries, retrieval prompting, category balancing, dedupe.
  • 450 question-SPARQL pairs across nine categories in the paper setting.

  RDFSPARQLbenchmarkingLLMs
  papergithub
• TrailKarma offline-first trail intelligence app with BLE relay and on-device ML

  Built during UCSD DataHacks: a delay-tolerant hiking app where phones pass trail reports through Bluetooth until someone reaches connectivity.

  AndroidBLEDatabricksoffline-first
  demogithub
• Multimodal Learning Analytics Dataset paper with EEG, eye tracking, GSR, video, and spatial reasoning data

  Work from IIT Bombay EdTech Lab on mental rotation tasks and learning analytics. I helped with multimodal data collection, validation, anonymization, and analysis.

  paperlearning analyticsmultimodal data
  paper
• VaultLens Telegram personal knowledge-base agent for memory, deadlines, digests, and calendar-aware context

  A recent build around the same thing I keep coming back to: making messy personal context useful without turning it into a dashboard chore.

  agentspersonal knowledgeTelegramcontext systems
  github
• Where Bits Matter in World Model Planning mixed-bit world-model planning study showing where low precision breaks

  A paired quantization study around DINO-WM. Same bit budget, different behavior: preserving precision in the right place changes planning.

  DINO-WMplanningmixed-bit
  papergithub