Suraj Ranganath

selected work

• C1 Vault post-quantum credential vault for AI agents, from scratch to launch

  Agents need access to systems without carrying the secrets that unlock them. I built C1 Vault from scratch through launch with a small core team, using Messaging Layer Security (MLS). I built egress proxies that inject credentials outside agent context, with scoped, short-lived access, rotation, and decoys to detect leaks. My work also covered the Tauri desktop app, CLI, and platform governance workflows, with support for zero trust deployment.

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

  Identity teams have to reason over messy access data: users, apps, roles, groups, exceptions, and stale permissions. If an AI system gets that wrong, it creates security risk instead of saving time. I helped build Copilot around the Identity Intelligence Graph, so natural-language questions could become permission-aware graph queries, reports, and guarded actions. It shipped into enterprise workflows, supported customer adoption and revenue, and became part of the patent work around agentic identity security.

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

  AI-text detection is too consequential to be a black-box score, especially when students and writers are the ones being judged. I worked on a detection interface that points to the specific spans behind a decision, so the output can be inspected instead of blindly trusted. That makes the detector easier to argue with, which is the whole point if these tools are going to be used around real people.

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

  AI-text detectors can fail in ways that matter for fairness, especially for people whose writing already gets treated as suspicious. I built StealthRL to stress-test that failure mode with reinforcement learning instead of one-off prompt tricks: train a paraphrase policy, preserve the meaning, and measure whether the attack transfers. The project turned a vague concern about detector robustness into a reproducible red-team system with a paper, demo, and public code.

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

  EEG researchers have years of workflows built around EEGLAB, but the center of gravity for scientific tooling has moved toward Python. I contributed to EEGPrep at the Swartz Center, helping port familiar EEGLAB-style workflows into a Python package without making researchers relearn everything from scratch. The work is open-source and practical: lower migration cost, better Python interoperability, and tooling that respects how EEG research is actually done. Arnaud Delorme's LinkedIn recommendation adds more context on my work there.

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

  Foundation-model training lives or dies on boring-seeming systems details: optimizer behavior, reproducibility, and clean integration with the training stack. Mentored by Al Merose, I implemented NAMO and NAMO-D support in Marin so researchers could try those optimizers inside a real open-source training codebase instead of a standalone experiment. The PR landed with tests and gave Marin first-class support for both optimizers.

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

  Long video generation is expensive because the model has to keep memory around while the sequence grows. I studied KV-cache compression for self-forcing autoregressive video generation across a broad set of methods, measuring whether the savings were worth the drift and quality tradeoffs. The paper gives a practical map of where low-bit cache policies can reduce memory pressure without quietly breaking the generated video.

  world modelsvideo diffusionquantizationsystems
  papergithub
• PIPE-Cypher automatic enterprise benchmark generation for Text-to-Cypher systems

  Text-to-Cypher systems are hard to evaluate in enterprise settings because public benchmarks do not resemble a company's own graph, data, or governance rules. I built PIPE-Cypher to turn a live property graph into an executable NL-to-Cypher benchmark while keeping private schema and values local. It turns benchmark creation from a hand-written chore into a repeatable evaluation pipeline, with thousands of accepted examples and downstream model comparisons in the paper.

  Text-to-Cypherproperty graphsbenchmarkslocal LLMs
  papergithub
• TrailKarma offline-first trail intelligence app with BLE relay and on-device ML

  Trails are exactly where useful local information matters and cell service disappears. At UCSD DataHacks, we built TrailKarma as an offline-first app where hikers can pass hazards, water reports, wildlife sightings, and delayed messages phone-to-phone until someone gets signal. I worked on the Android/BLE side and sync logic, and the project won hackathon awards because the demo solved a real no-signal problem.

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

  Standard grading misses a lot of student thinking, especially when the final answer is wrong but the reasoning is close. At IIT Bombay EdTech Lab, I worked on a multimodal learning-analytics dataset for mental rotation tasks, combining behavioral, physiological, interview, and task signals. The work became a Nature Scientific Data paper and shaped the education thread I keep returning to: using data to see how students think before reducing them to a score.

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

  Personal knowledge tools often fail because they ask for too much upkeep. VaultLens is my attempt at a lighter version: capture thoughts in Telegram, let an agent organize the context, and bring back the useful pieces when deadlines, reminders, or decisions come up. It is a small build, but it shows the kind of agent interface I care about: close to an existing habit and useful without becoming another dashboard.

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

  Efficient world models are only useful if compression does not destroy planning. In this study, I tested mixed-bit quantization around DINO-WM to see where precision matters most for spatial reasoning. The result was a clearer picture of the transition zone: some low-bit settings stay close to full precision, while the wrong allocation can break planning even at the same overall budget.

  DINO-WMplanningmixed-bit
  papergithub
• PIPE-RDF LLM-assisted pipeline for enterprise RDF benchmark generation

  Enterprise knowledge-graph QA has the same evaluation problem across formats: public datasets rarely match the schema people actually use. PIPE-RDF tackles that for RDF/SPARQL by generating schema-specific question-query pairs and checking that they execute. That makes the benchmark useful for real KGQA evaluation, not just another cleaned-up public toy dataset.

  RDFSPARQLbenchmarkingLLMs
  papergithub