01.About

Based in Bangalore, India

02.Experience

1. Jan 2024 — Present

   Founder · Tether AI

   Bangalore, India

   • Building Tether AI, an early-stage product: a runtime safety layer for local AI agents that checks every action an agent takes before it runs.
   • Core features: a policy engine, taint tracking, runaway-agent protection and human-in-the-loop approvals, with a full audit trail.
   • Raised pre-seed funding from an investor through the college incubation center, Big Foundation (BMSCE); currently in the product-building stage.
   • It stops risky agent actions with 92.9% policy accuracy while adding under 1 ms of latency, so teams get protection without slowing their agents down.

   AI clients supported out of the box

   5

   command to install and protect an agent

   1

   policy enforcement accuracy

   92.9%

   added latency per agent action

   <1 ms

   • Node.js
   • MCP
   • AI Agent Security
   • Policy Engine

2. Nov 2025 — Jan 2026

   Research Intern · Council of Scientific and Industrial Research (CSIR)

   • Author of the paper “Regime-Aware Cooperation in Multi-Echelon Inventory Control: Curriculum Pretraining and a Zero-Shot LLM Supervisor”, which turns the MARL cooperation weight α from a fixed hyperparameter into a run-time control input.
   • Trained a single HAPPO policy (GRU actors, centralized critic) across three cooperation regimes, α ∈ {0.100, 0.375, 0.650}, under a 70/15/15 skewed curriculum for 3×10⁶ steps on a four-echelon serial supply chain with Merton jump-diffusion demand.
   • Cut per-period cost by 12.8% under a pulse demand shock and 6.9% under a demand collapse versus a reproduced single-α specialist, with α frozen at execution time, so the gain needs no supervisor and no API call at run time.
   • Raised in-distribution customer fill rate by +7.1 points (55.7% → 62.8%) and service level from 31.0% → 43.3% at cost parity (p ≈ 10⁻¹²).
   • Built a dual-timescale LLM supervisor that reassigns a per-echelon α vector every 20 periods from window telemetry. Zero-shot, it matched a hand-tuned rule base to within ±3 cost units (n = 100) with no environment-specific thresholds and an auditable rationale for every decision.
   • Benchmarked on 280 frozen traces (56,000 periods) spanning in-distribution, pulse, step-up and collapse demand, using paired t-tests with Holm correction; the pulse and collapse cost gains survive correction.

   cost under a pulse demand shock

   −12.8%

   cost under a demand collapse

   −6.9%

   customer fill rate, at equal cost

   +7.1 pts

   cost-unit gap: zero-shot LLM vs hand-tuned rules

   ±3

   • Multi-Agent RL
   • HAPPO
   • Curriculum Learning
   • GRU
   • LLMs
   • Python

3. Mar 2024 — Apr 2024

   Intern · Sankara Eye Hospital

   Bangalore, India

   • Mapped the OPD (outpatient) patient journey end to end, from registration through consultation to cross-department referral, and analyzed it with Python to find bottlenecks, high-delay stages and referral leakage (referred patients who drop out before completing the referral).
   • Defined and tracked 8 operational KPIs used in hospital OPD management: OPD wait time, stage-wise waiting time (registration → pre-consultation → consultation), visit cycle time, patient throughput, appointment no-show rate, referral turnaround time (TAT), referral leakage rate and doctor utilization. Standardized the preprocessing and reporting pipelines behind them so progress and exceptions were visible to non-technical stakeholders.
   • Built KPI dashboards that surfaced the highest-delay stages so the operations team could prioritize fixes.
   • Recommendations from the analysis reduced cross-referral wait times by 30%.

   cross-referral wait time

   −30%

   operational KPIs defined and tracked

   8

   • Python
   • Data Analysis
   • Process Mapping
   • KPI Dashboards

03.Projects

• unhalted: AI Revenue Recovery Agent★ featured

  A mandate-recovery agent for Razorpay subscriptions. When a mandate debit fails, the standard flow retries blindly until the subscription is halted; unhalted diagnoses why the debit failed and routes each case to the right recovery path. The model diagnoses; a deterministic shell decides.

  • Architecture: signature-verified webhook ingest → diagnosis via a rules table over Razorpay's error_reason / error_source (model only for ambiguous cases) → 0.70 confidence gate → deterministic shell for windows, caps, stops, escalation ladder, compliance lint and payment links → leased, at-least-once, idempotent action queue on SQLite with an append-only audit trail.
  • Closes the loop: recovery payment links carry the case ID, so a payment_link.paid webhook marks the case RECOVERED and cancels any pending actions.
  • Compliance enforced in code, never by the model: NPCI execution windows, the 25-hour pre-debit rule, the ₹15,000 UPI Autopay ceiling, mandate max_amount, a retry cap of 3, contact hours, one message per customer per fortnight and 9 hard stops (revocation, opt-out, dispute, distress, chargeback, regulatory hold and more).
  • Benchmarked against Razorpay's documented retry policy on 300 failures (₹1,20,850 at risk): 84 vs 900 debit attempts, 0 vs 108 attempts a retry cannot fix, 0 vs 117 attempts in NPCI restricted bands and 189 targeted customer contacts where the baseline sends none. 27 ambiguous cases are held for a human instead of guessed.
  • LLM reply parsing for Hinglish and English customer replies: 68 / 68 parsed, 1.00 recall on opt-out, distress and promise-to-pay, at $0.00007 per parse and a 1.0 s median latency.

  debit attempts (84 vs 900)

  −91%

  retries that provably can't succeed

  0 vs 108

  diagnoses needing a model call

  0 / 300

  automated tests

  538

  • Python
  • Razorpay
  • LLMs (OpenRouter)
  • SQLite
  • ASGI / uvicorn
  • pytest
  • Gherkin

• EUDR Evidence Agent for Karnataka Coffee Farmers

  The EU Deforestation Regulation (EUDR, Regulation (EU) 2023/1115) allows coffee, cocoa, palm oil, soy, cattle, rubber and wood onto the EU market only if they were legally produced on land not deforested after 31 December 2020, proven with the geolocation of every plot. After two postponements, it applies from 30 December 2026 for large and medium operators and 30 June 2027 for micro and small enterprises. For Karnataka's shade-grown coffee in Kodagu, Chikmagalur and Hassan, every plot will need that evidence before its coffee can be sold into the EU.

  • What it does: answers "can coffee from this plot go into an EU-bound lot, and if not, what evidence is missing?" with a Clear / Needs evidence / High risk verdict, a cited evidence file and a list of the documents still missing.
  • Architecture: a LangGraph supervisor agent (ReAct, capped at 10 steps) delegates to a geo-analyst agent for satellite checks and a rules agent for legal lookup. A reviewer checks that every conflict is explained, a fixed verdict table makes the call, and an officer approval pause gates every output.
  • Built the supervisor and human-approval layer: geometry validation against EUDR geolocation rules, deterministic verdict rules, checkpointed pause and resume, Langfuse tracing of every step and tool call, and a per-plot result cache.
  • What each dataset proves: Hansen GFC, tree-cover loss after 2020 by year; JRC TMF, whether that loss was clearing or degradation; JRC GFC2020, the EU's own forest map at the 31 Dec 2020 cut-off; FDaP coffee, whether the land was already farmed as coffee in 2020; Sentinel-2, dry-season greenness to tell shade-tree pruning from clearing; WDPA, overlap with protected areas.
  • Legal evidence: the rules agent searches EUDR 2023/1115, 2025/2650 and Commission guidance, grades the retrieved passages, rewrites weak queries (up to 2 retries) and must cite the exact article, or report that no rule was found.
  • Evaluation plan: ~40 labelled test plots, measuring verdict accuracy, missed and false flags, citation correctness and cost, plus an A/B test against Whisp's label alone.
  • Python
  • LangGraph
  • Google Earth Engine
  • Whisp
  • Chroma / FAISS
  • Langfuse
  • Sentinel-2

• Cloud-Native Static Site Deployment Platform (Mini Vercel)★ featured

  A platform that builds, packages and hosts static web apps from GitHub repositories or uploaded archives in isolated Docker build environments.

  • Asynchronous deployment pipeline with FIFO worker queues, retries, versioning and immutable AWS S3 artifacts for fault-tolerant delivery.
  • EC2 workers clone repositories, install dependencies, run custom builds and publish deployments with runtime isolation.
  • React dashboard with authentication, deployment history, build logs, live status and Nginx custom-subdomain routing, plus CI/CD and cloud monitoring.
  • React
  • FastAPI
  • Docker
  • AWS EC2
  • AWS S3
  • Nginx
  • PostgreSQL
  • Redis
  • Bash
  • GitHub Actions

04.Research Papers

• Regime-Aware Cooperation in Multi-Echelon Inventory Control: Curriculum Pretraining and a Zero-Shot LLM Supervisor

  Multi-agent RL supply-chain policies usually fix the cooperation weight α before deployment. We make α a run-time control input, set by one HAPPO policy trained across three cooperation regimes and a supervisor that reassigns it every 20 periods.

  • Curriculum pretraining alone delivers the robustness, with no supervisor and no API call at run time.
  • A zero-shot LLM supervisor matches hand-tuned rules with no environment-specific thresholds, evaluated on 280 frozen traces.
  • Multi-Agent RL
  • HAPPO
  • Curriculum Learning
  • LLM Supervisors
  • Supply Chain

• Agentic Context Engineering and Self-Learning Large Language Models: A Comprehensive Review, Taxonomy, Architectures, and Open Research Challenges

  A PRISMA-guided systematic review of how LLM agents learn after deployment without updating weights, by curating the model's context as persistent system state.

  • Contributed to data curation, formal analysis and the original draft.
  • Introduces a mechanism/system distinction and a 7-dimension design space that places systems from RAG and Reflexion to ACE.
  • Finds that feedback reliability matters more than hyperparameters, and that rewriting the whole context is a documented failure mode.
  • LLM Agents
  • Context Engineering
  • Self-Learning LLMs
  • Systematic Review
  • Transformers

• Agentic Context Engineering for Self-Learning LLM Agents: A Critical Review of Mechanisms, Evidence, and Risks

  A critical review of LLM agents that improve after deployment by rewriting what they read instead of their weights, with a quantitative synthesis across systems and a security threat model.

  • Contributed to data curation, formal analysis and the original draft, and co-checked every inclusion decision (97.9% agreement).
  • Coded 22 core systems against a six-function adaptation loop and a seven-dimension design space.
  • Showed that feedback reliability, not store design, decides outcomes: median gains of +8.3 pts with labels vs +1.1 pts with self-assessment, where 17 of 45 results fell below baseline.
  • Mapped 11 attacks and 6 defenses for adaptive agent memory.
  • LLM Agents
  • Context Engineering
  • Agent Memory
  • Critical Review
  • AI Security

05.Open Source

• JAX

  Working on Pallas-based TPU kernel development.

• Kubeflow, Keras, MLflow (working on)

  Working on open-source issues in ML infrastructure and deep-learning frameworks.

06.Skills

Programming

• C++
• Python
• Java
• TypeScript
• SQL
• Rust
• Bash

Computer Science

• Data Structures & Algorithms
• Object-Oriented Design
• Operating Systems
• Computer Networks
• DBMS
• Complexity Analysis
• System Design

Backend Systems

• Python backend (FastAPI)
• Go backend
• Microservices
• REST APIs
• Message Queues (RabbitMQ, Kafka)
• Event-Driven Architecture
• Concurrency
• Caching

Databases

• PostgreSQL
• MongoDB
• Redis
• SQL
• pgvector

Cloud & Infra

• AWS (EC2, S3)
• Docker
• Kubernetes
• Terraform
• Linux
• Nginx
• CI/CD (GitHub Actions)
• Prometheus
• OpenTelemetry
• Git/GitHub

AI/ML

• Machine Learning
• Deep Learning
• Computer Vision
• Reinforcement Learning (MARL)
• Recommendation Systems
• LLMs
• RAG
• Agentic AI
• LLM Guardrails
• AI Automation Pipelines
• MLOps (MLflow, Kubeflow)
• Distributed ML
• Model Evaluation

07.Education

1. 2023 — 2027

   B.E. in Computer Science (Data Science)

   BMS College of Engineering

   CGPA 9.14 · Expected graduation 2027

   Relevant coursework

   CS core

   • Data Structures
   • Algorithms
   • Operating Systems
   • Databases
   • Computer Networks

   Specialisation

   • Machine Learning
   • Deep Learning
   • Computer Vision
   • Distributed Computing
   • Agentic AI

08.Achievements

• ▹

  Finalist, JPMC Code for Good

• ▹

  Winner, Startup Pitching Event

  Pitched Tether AI to judges and peers, earning recognition for innovation, feasibility and growth potential.

• ▹

  President, Big Foundation

  The incubation center of BMSCE.

• ▹

  NSS Volunteer, BMSCE

  Organized cleanliness and awareness drives and coordinated student participation in community outreach programs.