Overview
Who I am, and the record behind every line on this page
AI security engineer working on the failure mode where an autonomous agent does something a human would have blocked. A patch of mine is merged into google/go-github; five memory-safety issues I filed against google/s2geometry are open, including a 2.4 GiB allocation reachable from a 28-byte input, with standalone reproducers for each. Two papers accepted — one on local policy verification for agents (ICICSET 2026), one on predictive bandwidth scaling for WebRTC over LEO satellite links (IEEE RTC 2026, whose research track is published in IEEE Xplore). Ranked #1 on the HackingHub Q3 2026 global leaderboard (177 flags, 18,125 XP). Two years as a full-time AI engineer at Atmos SoftTech, building production ML systems, full-stack products and offline-first agent pipelines.
I would rather be checked than believed: the claims about my own technical work that the ledger covers are re-checked weekly against their public source by a script anyone can run — github.com/sushant-me/reputation.
Auditing six AI-action rules in sisaku-security/sisakulint surfaced defects in five. Instead of taking the premise on trust, maintainer on-keyday built both revisions of the rule and ran them over the same inputs, then wrote "the gap is real and the direction you took to it is a reasonable one" and filed three tracked issues from the audit. The pull request was closed for review capacity, not for being wrong. The quotation is machine-checked in my reputation repository — edit that comment and the claim goes red.
177 flags and 18,125 XP, with 1 gold, 2 silver and 2 bronze awards; the next-ranked account holds 97 flags. Public leaderboard API.
My PR #4556 stopped a release-asset upload being sent to an off-host URL. The maintainer merged it, then replaced it the next day with the broader PR #4564 — "credentials are sent only to configured origins" — which is what master implements today and which carries the commit "Address feedback from sushant-me". My name is in the project's AUTHORS file on master.
A NULL dereference on the library's documented traversal path, an out-of-bounds read, a 16 GiB allocation from a 117-byte input, and a 2.4 GiB allocation reachable from a 28-byte input. Two hardening pull requests are under review by the maintainer. The maintainer's reply to these, quoted rather than paraphrased: "It's not really a DoS since these are assumed to only operate on trusted data. We should document the assumptions better." — a disagreement about severity and threat model, not about whether the reads are out of bounds.
google/libphonenumber's as-you-type formatter asserted 0.00% line coverage of the code it targeted because it never generated valid input; the corrected harness reaches 93%.
Recognized consecutive years (2024 & 2025) for social entrepreneurship and business models addressing global challenges.
Selected as one of roughly 10,000 finalists from 54,000+ applicants worldwide, completing leadership training co-created by Harvard Business School faculty.
Logistics, executive administration and student engagement for department-level technical events and hackathons at Nepal Engineering College.
Poster on 5G non-terrestrial-network connectivity from ordinary smartphones to LEO satellites, sized for Nepal's terrain.
Experience
Two years full-time, and the roles around it
Two years full-time designing, developing and deploying AI-based systems and machine-learning models in production. Built and maintained full-stack web applications end to end, and engineered autonomous agent pipelines — tool-calling, multi-step task orchestration, retrieval over private local data, and offline-first inference where sending data to a cloud API was not acceptable. Conduct and performance rated highly satisfactory in the company's signed experience letter.
Capstone at an AI studio working on agent runtimes and human-in-the-command architectures for emerging Asian markets, building autonomous agent pipelines with RAG and LangChain.
Core full-stack technical lead architecting corporate web ecosystems and mobile applications. Managed enterprise IT infrastructure for 30+ computer setups. Engineered automated inventory management with hardware-software integration. Spearheaded digital marketing initiatives and brand visibility campaigns.
Leading cybersecurity company in Nepal since 2013. Conducted penetration testing, vulnerability assessments, and security audits for enterprise clients across Asia, Europe, Australia, and the USA. Discovered critical CVE vulnerabilities and delivered structured remediation reports.
Delivered foundational AI and computer science education through structured coaching programs. Developed curriculum modules for emerging technology literacy, practical programming skills, and hands-on machine learning workshops for students.
Research
Accepted papers, and work still in draft
Accepted. Predictive bandwidth scaling for real-time WebRTC video over low-Earth-orbit links, where round-trip times and handover gaps break conventional congestion control. A lightweight neural predictor scales the send rate ahead of the link degradation instead of reacting to it, evaluated against reactive baselines over emulated LEO conditions.
Accepted. A structured Chain-of-Thought verification framework that asks whether a small, locally served model can enforce policy for an agent that executes actions rather than proposing them. 600 policy scenarios per condition on Phi-3-mini (3.8B, 4-bit) on one consumer laptop inside a 4.2 GiB VRAM budget, with no cloud inference. The result contradicted our hypothesis: constraining output to JSON without a reasoning field was the least safe of the four arms — it approved 46.2% of the 600 proposals it was shown, against 17.2% for unconstrained text, while a mandated reasoning field cut false accepts of irreversible hard-denial commands from 71 to 6. The residual failures are reported rather than rounded away.
A proof of concept in digital forensics: whether an open knowledge format can carry incident data between cooperating agents so each traverses only the part of an incident graph it needs, rather than ingesting the whole corpus. Co-authored with Rakhee Pandey.
Fixes and reports for framework-reserved tool names that a server-supplied MCP tool can occupy, bypassing the duplicate-name guard: an in-model built-in such as google_search, or the terminal tool set_model_response. Reproduced in Go and Java against the real frameworks, with the reserved sets corrected against each port rather than copied between them.
Engineered a fully offline robotic OS eliminating cloud dependency. Utilizes C++ hardware control and local LLMs for manipulation tasks, with RAG-based episodic memory for continuous learning. Everything runs with no network connection at all.
Co-authored with Aashika Pandey, who is the first author. Proposed the Deep Structural Causal Model (DSCM) architecture to systemically resolve AI hallucinations by mathematically embedding causal graphs into generative networks. Integrates Pearl's do-calculus for robust counterfactual reasoning.
Comprehensive architecture for universal 5G Non-Terrestrial Network connectivity connecting smartphones directly to LEO satellites, tailored for Nepal's topography. Targets under 50ms latency for universal connectivity in remote regions.
Co-authored with Sulav Timalsina, who is the first author. Strategic evaluation of Nepal's high-altitude climate and hydropower for sustainable South Asian data infrastructure, analysing the topological, climatic and energy-infrastructure case.
A fully offline, sovereign AI governance stack: a quantized local model for assistance behind a deterministic gatekeeper that blocks data-extraction attempts. Presented at the Municipal AI Governance Conference 2026.
Conducted offensive security research on covert surveillance via Wi-Fi Channel State Information (CSI). Proposed deterministic mitigation protocols for enterprise IoT environments against Wi-Fi sensing-based attacks.
Architected a theoretical framework for bounding and actively verifying LLM-driven autonomous agent decisions in real-time. Effectively mitigates unsafe state transitions in multi-agent collaborative environments.
Projects
Shipped tools, upstream patches, and published advisories
Mamba's S6 recurrence from scratch, measured honestly
Final Capstone — Tourism App
AI Sentiment & Trust Engine
Offline Payment Wallet — Hackathon
Nepal Engineering College Utility
Logistics & Ride-Hailing Engines
Hardware-Software Integration
NLP-Powered Opinion Mining
Cross-Platform Media Bridge
Dynamic Car & Bike Pricing
Real-Time Order Tracking
Encryption & Security Utilities
Syllabus & Academic Navigator
A glob deny rule could be stepped around with a newline in the matched value, so the call fell through to a broader allow and proceeded with no human in the loop. The matched value is attacker-influenced: tool names come from the MCP server being gated.
Replies were read without a bound, so a server returning one enormous body or line could exhaust memory and kill the scan. The tool exists to inspect servers it does not trust, so dying on the reply fails open against the exact adversary it was pointed at.
The masker blanked any string that was the whole right-hand side of an assignment, so a one-line message constant became empty and a duplicate-registration rule reported nothing. A false negative in a detector, which its own documentation calls worse than the false positive being fixed.
The invisible-character ranges covered U+FE00..U+FE0F and called it "variation selectors". The supplement is the same channel, and a payload hidden in it survived both the scan and the helper meant to show a reviewer what the text really contains.
Achievements
Awards, honours, and the moments behind them
Certificates
Every certificate below opens its own PDF
Skills
What I actually work with, day to day