Rishab Satpathy@rishabsatpathy146
2
Events attended
2
Submissions made
India
1 year of experience
About me
I'm an 18-year-old B.Tech engineering student at ABV-IIITM Gwalior focused on the intersection of low-level systems and high-reliability AI. My work spans the full computing stack: from publishing research on the physical constraints of non-Von Neumann CPU interconnects, to building applied AI systems like LISTEN (a predictive maintenance model using autoencoders/LSTMs that was a Hacksagon 2026 finalist) and physiological EEG signal processing models. Currently, I am co-architecting TAPES—a deterministic AI compiler that uses AST-level physics to eliminate LLM hallucinations in autonomous code generation. I believe in event-driven execution and building mathematically bounded systems that don't just work, but structurally refuse to fail.
🤝 Top Collaborators
🤓 Latest Submissions
TAPES(Transition AI Patches Enforcement System)
TAPES: Deterministic Governance for Autonomous Enterprise AI The Problem: The Risk of Agentic Autonomy Autonomous AI agents like IBM Bob excel at reasoning and patch generation, but they lack physical boundary awareness. Without strict oversight, an autonomous agent can accidentally overwrite legacy modules, bypass security invariants, or introduce catastrophic structural regressions into mission-critical codebases. The Solution: Runtime Governance TAPES is a deterministic security layer engineered specifically to secure IBM Bob and the watsonx ecosystem. It acts as an absolute runtime gatekeeper, intercepting every code modification payload generated by the IBM Bob API before it touches the local filesystem. How It Works: The Five-Gate Bouncer TAPES routes all LLM outputs through three decoupled execution layers (kernel, runtime, execution), guarded by our custom AST-level Bouncer. The Bouncer validates all patches against strict physical boundaries: Scope Integrity: Prevents "ghost" variable references. Dependency Isolation: Intercepts unauthorized external imports. Structural Containment: Blocks malicious path-traversal attempts. File Protection: Enforces write-locks on sensitive architecture. Empirical Performance & Cost Efficiency In our automated 8-scenario benchmark using live IBM Granite models, TAPES proved its enterprise viability: 100% Production Safety: Unprotected BobShell executions compromised protected files 8/8 times. TAPES successfully intercepted 26 unsafe modifications, achieving a 100% containment rate. 81% Token Reduction: By utilizing retrieval shaping and strict context bounding, TAPES slashed IBM Granite token consumption from 2.91 million to just 530,000—delivering massive compute cost savings while allowing legitimate patches to flow seamlessly. Business Value TAPES transforms unpredictable agentic workflows into secure, cost-effective, and deterministic development pipelines ready for the enterprise.
19 May 2026
TAPES (Transition Ai Patches Enforcement Systems)
Most AI coding agents are just LLMs with write access. They hallucinate symbols, overwrite protected files, and flood your codebase with untested patches. TAPES fixes that. TAPES (Transition Ai Patches Enforcement Systems) wraps IBM BobShell — Granite-13b-chat-v2 via watsonx.ai — in a five-stage governed execution pipeline that transforms reckless code generation into controlled, auditable engineering. The pipeline starts before Bob sees a single line of code. An AST-powered Intent Router computes a centrality-weighted token budget and extracts scope directives from the user's intent. A frozen Contract locks in the mutation boundary, constraints, and success criteria. Bob only ever sees hollow skeletons of irrelevant files and the full source of the exact target symbol. That single design decision cuts token consumption by 81%. Bob's patches hit a five-gate Bouncer before touching the filesystem: syntax validation, path boundary enforcement, AST anchor verification, exact-match uniqueness, and size integrity. Protected files are never hard-blocked — patches targeting them redirect silently to sidecar extension files, preserving intent without violating protection invariants. Surviving patches enter a dual-tier sandbox. Tier 1 runs targeted tests against a fast Python-only copy in milliseconds. Tier 2 clones the full virtual environment for final isolation before atomic commit. Failures are classified as transient, structural, or semantic — each carrying a different retry penalty — and prior failure tracebacks are injected directly into Bob's next prompt. Across 8 adversarial benchmarks — protected file massacres, mutation floods, path traversal attacks, hallucinated symbols — raw BobShell violated protected files 8/8 times and cost $7.29. TAPES delivered 0 violations, $1.34 total, 26 Bouncer intercepts, and 6 legitimate patches committed cleanly. Bob writes the code. TAPES decides if it deserves to exist.
17 May 2026
