Founder-engineer building AI security, agent evaluation, multi-agent systems and neuro-symbolic products
I am a computer science engineering student and founder of TRUSCOR, working at the intersection of AI security, adversarial evaluation, interpretable evidence, multi-agent systems, reinforcement learning and applied deep tech.
My work usually starts with an ambitious system-level question. How do we reproduce an agent failure? How do we separate perception from reasoning? How do we make autonomous software auditable? I turn those questions into architectures, working prototypes, evaluation environments, developer tools and open-source infrastructure.
CIBIL scores borrowers and never lends a rupee, which is why every bank trusts it. We are the bureau, not the bank.
TRUSCOR builds no models, no agents and no applications. It tests what AI systems do, records it as signed evidence, and rates how far they can be trusted. The organisation lives at github.com/TRUSCOR.
| Layer | What it does | Access |
|---|---|---|
| SOVA OSS | Map, test, record, replay and investigate AI agents. The open .sova evidence format. |
Open source, Apache-2.0 |
| SOVA Engine | Attacks and stress-tests AI products through the interface layer, into .sova evidence |
Rating team, licensed firms |
| SOVA Black Box | A 24x7 recorder for AI running in production | Deployments, OEM partners |
| O.W.L | The measurement instrument: checks what an AI claimed against what really happened | Internal, licensed measurers |
| TAFAAR | Grades evidence into the TRUSCOR Rating (300-900) and an Estimated Financial Exposure | TRUSCOR only, never licensed |
No model, API key or network needed:
git clone https://github.com/TRUSCOR/SOVA-OSS
cd SOVA-OSS
uv sync --locked
uv run sova demo sleeperLatest preprint: .sova: Portable, Evidence-Bounded Reproduction of AI-Agent Security Findings (version 1.1, November 2026, not yet peer reviewed).
I re-ran a widely used AI-agent security benchmark on eight models. It can score
a model as attacked when the model warns its user about the attack, and attacks
published in 2024 still worked on GPT-4o in 2026. Every finding ships as a
.sova file that anyone can verify offline.
| Where | What |
|---|---|
| TRUSCOR/research | Released papers, with data, code, protocols and .sova evidence |
| research-papers | All my preprints and working drafts, across SOVA, TRUSCOR and other projects |
| SOVA OSS v1 | The original SOVA prototype, started 28 July 2026, preserved and frozen |
| Project | What I built | State |
|---|---|---|
| SOVA OSS | A local-first, open-source instrument for recording, packaging, replaying, reproducing and comparing AI behaviour. Portable behaviour artifacts, signed traces, deterministic oracles, containment controls, capability mapping and the foundations of an agent-forensics engine. | Active pre-alpha |
| Bhai | A proactive personal AI agent in TypeScript and Rust for computer use, deep research and long-running tasks across Telegram, WhatsApp, Slack, Discord and desktop. | Source-available |
| SOVA Engine | The TRUSCOR AI red-team engine. Coordinates multi-agent reconnaissance, attack, evaluation, judging and evidence workflows across browser, UI and tool surfaces. | Public showcase |
| Neural Forge | An end-to-end specialist-model pipeline: data synthesis, LoRA/QLoRA supervised fine-tuning, GRPO-style alignment experiments, evaluation and GGUF export. | Functional prototype |
| SOVA Research | Released papers on portable, verifiable evidence for AI-agent security findings, with their data, code, protocols and evidence as .sova files. |
Preprint v1.1 released |
| ORACLE Prompt | A versioned system-prompt architecture for cognitive triage, adversarial hypothesis generation, epistemic honesty and strategic synthesis. | Public, documented evolution |
| Repository | What it is | Stage |
|---|---|---|
| SOVA OSS | Portable AI-behaviour records, signed evidence traces, offline verification, replay, trigger search, containment, deterministic oracles, agent forensics | Active pre-alpha |
| SOVA Research | Released papers behind the .sova capsule and .sova-trace record, with data and .sova evidence |
Preprint v1.1 |
| Research Papers | All my preprints and working drafts, across SOVA, TRUSCOR and other projects | Preprints and drafts |
| SOVA OSS v1 | The original SOVA prototype, started 28 July 2026, preserved with its history | Frozen |
| NOVA | Synthetic enterprise-finance workspace and intentionally vulnerable AI-security test target for prompt-injection, data-access and policy-bypass evaluation | Implemented testbed |
| Hawkeye | An OpenEnv-compatible RL environment where agents audit vehicle-crash liability by simulating collisions and checking witness testimony against physics | RL environment |
| Repository | What it is | Stage |
|---|---|---|
| Bhai | Proactive personal AI agent, TypeScript + Rust, computer use and long-running tasks across five messaging surfaces and desktop | Source-available |
| Neural Forge | Specialist-model pipeline: synthesis, LoRA/QLoRA, GRPO experiments, evaluation, GGUF export | Functional prototype |
| Viber AI | Multi-agent software-building platform turning text or voice ideas into requirements, code, review workflows and downloadable projects through a live IDE | Product prototype |
| Oracle_101 | Full-stack AI education platform: tutoring, assessments, OCR, translation, voice/TTS, learning analytics, spaced repetition | Large working codebase |
| ORACLE Prompt | Versioned system-prompt architecture for triage, adversarial hypothesis generation and epistemic honesty | Public prompt system |
| Repository | What it is | Stage |
|---|---|---|
| Robodog | PPO locomotion for a Unitree quadruped in Genesis: walking, running, yaw control, reward shaping, proprioceptive observations, sim-to-real considerations | RL training experiment |
| Agro-X | AI agriculture assistant: crop-image analysis, soil and weather context, recommendations, market guidance, voice access, SMS workflows | Product prototype |
System-Designs collects six architecture specifications, each imported with its original commit history: Veltrix (energy-fingerprint security), Watson (forensic reconstruction), NeuroPilot (developer copilot), Vacay (travel concierge), Synaptek (ICU early warning) and Axon (structural inspection).
These are design specifications, not implementations, and they say so.
I am especially interested in collaborations around:
- agentic AI security, red teaming and behavioural evaluation;
- reproducible AI evidence, incident reconstruction and interpretability;
- multi-agent orchestration and developer tools;
- reinforcement learning environments and neuro-symbolic systems;
- ambitious 0 to 1 deep-tech products.
"What I cannot create, I do not understand."
— Richard Feynman
See the system. Test the behavior. Preserve the evidence.